Pediatric and adult congenital cardiac phenotype prediction using electrocardiogram

Abstract

Disclosed herein are techniques for evaluating cardiac health of a pediatric patient or adult congenital patient, including techniques for generating one or more models (e.g., machine learning models) trained to predict a cardiac abnormality in a pediatric patient or adult congenital patient. Some methods may include receiving electrocardiogram (ECG) data of a pediatric patient or adult congenital patient, determining whether the patient corresponds to one or more of multiple cardiac phenotypes, each corresponding to a presence of a cardiac abnormality based on analysis by one or more trained models, and outputting the cardiac phenotype(s) based on the ECG data. In some embodiments, the cardiac abnormality may include one or more of mortality, ventricular dysfunction, ventricular dilation, ventricular hypertrophy, one or more rhythm abnormality, one or more ECG abnormality, and critical congenital heart disease.

Description

PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025PEDIATRIC AND ADULT CONGENITAL CARDIAC PHENOTYPE PREDICTION USING ELECTROCARDIOGRAMCROSS REFERENCE TO RELATED APPLICATIONS

[0001] The present application claims the benefit under 35 U.S.C. § 119(e) of U.S. Provisional Patent Application No. 63 / 701,312, filed September 30, 2024, and titled “Pediatric And Adult Congenital Cardiac Phenotype Prediction Using Electrocardiogram,” the entire contents of which are incorporated herein by reference.BACKGROUND

[0002] Pediatric and adult congenital patients, both healthy and those experiencing cardiac difficulties, may need to be evaluated for whether their symptoms demonstrate the presence of a particular cardiac-related medical condition. Patients have generally relied upon MRI or echocardiograms for cardiac screening, which have high reliability and have been in wide use.SUMMARY

[0003] In some aspects, the techniques described herein relate to a method of predicting a cardiac abnormality in a pediatric patient or adult congenital patient, the method including: determining, based on received electrocardiogram (ECG) data of a pediatric patient or adult congenital patient, one or more of a plurality of cardiac phenotypes to which to assign the pediatric patient or adult congenital patient, at least one cardiac phenotype of the plurality of cardiac phenotypes each corresponding to presence of a cardiac abnormality, the determining the one or more of the plurality of cardiac phenotypes to assign to the pediatric patient or adult congenital patient including analyzing the received ECG data using one or more trained models, wherein the one or more trained models were trained with training data including ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert; and outputting the one or more of the plurality of cardiac phenotypes determined for the pediatric patient or adult congenital patient based on the ECG data.

[0004] In some aspects, the techniques described herein relate to a method, wherein determining the one or more of a plurality of cardiac phenotypes to which to assign the pediatric patient or adult congenital patient includes determining a probability of the pediatric patient or adult congenital patient having one or more cardiac abnormalities.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0005] In some aspects, the techniques described herein relate to a method, wherein: determining a probability of the pediatric patient or adult congenital patient having one or more cardiac abnormalities includes determining a probability of the pediatric patient or adult congenital patient having any of the one or more cardiac abnormalities; and determining the one or more of the plurality of cardiac phenotypes to which to assign the pediatric patient or adult congenital patient includes, based on the probability of the pediatric patient or adult congenital patient having any of the one or more cardiac abnormalities, determining whether to assign the pediatric patient or adult congenital patient to a phenotype for pediatric patient or adult congenital patients with any of the one or more cardiac abnormalities.

[0006] In some aspects, the techniques described herein relate to a method, wherein determining a probability of the pediatric patient or adult congenital patient having one or more cardiac abnormalities includes determining a probability of the pediatric patient or adult congenital patient having each of the one or more cardiac abnormalities, wherein the one or more cardiac abnormalities include mortality, ventricular dysfunction, ventricular dilation, ventricular hypertrophy, one or more rhythm abnormality, or critical congenital heart disease.

[0007] In some aspects, the techniques described herein relate to a method, wherein determining the one or more of the plurality of cardiac phenotypes includes comparing the probability of the pediatric patient or adult congenital patient having each of the one or more cardiac abnormalities to a threshold.

[0008] In some aspects, the techniques described herein relate to a method, wherein determining one or more of the plurality of cardiac phenotypes includes determining one or more of a mortality risk, ventricular dysfunction phenotype, ventricular dilation phenotype, ventricular hypertrophy phenotype, one or more rhythm abnormality, or critical congenital heart disease.

[0009] In some aspects, the techniques described herein relate to a method, wherein determining one or more of the plurality of cardiac phenotypes includes determining whether the pediatric patient or adult congenital patient has a ventricular dysfunction.

[0010] In some aspects, the techniques described herein relate to a method, wherein determining one or more of the plurality of cardiac phenotypes includes determining whether the pediatric patient or adult congenital patient has ventricular dilation.

[0011] In some aspects, the techniques described herein relate to a method, wherein determining one or more of the plurality of cardiac phenotypes includes determining whether the pediatric patient or adult congenital patient has ventricular hypertrophy.

[0012] In some aspects, the techniques described herein relate to a method, wherein determining one or more of the plurality of cardiac phenotypes includes determining whether the pediatric patient or adult congenital patient has a risk of mortality.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0013] In some aspects, the techniques described herein relate to a method, wherein determining one or more of the plurality of cardiac phenotypes further includes assigning a qualitative assessment for the one or more of the plurality of cardiac phenotypes, wherein the qualitative assessment indicates whether a severity of a cardiac abnormality indicated by a cardiac phenotype.

[0014] In some aspects, the techniques described herein relate to a method, wherein determining the one or more of the plurality of cardiac phenotypes based on the received ECG data includes determining the one or more of the plurality of cardiac phenotypes based on one or more of: at least one ECG raw waveform, a QRS interval, a QRS axis, a T axis, a P axis, a PR interval, a QT interval, a QT corrected for heart rate (QTc), or a heart rate.

[0015] In some aspects, the techniques described herein relate to a method, wherein determining using the one or more trained models includes determining using the one or more trained models were trained with training data that did not include data for pediatric patients having congenital heart disease.

[0016] In some aspects, the techniques described herein relate to a method, wherein determining using the one or more trained models includes determining using the one or more trained models were trained with training data including data for pediatric patient or adult patients having congenital heart disease.

[0017] In some aspects, the techniques described herein relate to a method, wherein determining using the one or more trained models includes determining using one or more trained models were trained with training data for patients across a plurality of age ranges, each of the plurality of age ranges being a range within an overall age range of 0 to 18 years, and greater than 18 years for adult congenital heart disease.

[0018] In some aspects, the techniques described herein relate to a method, wherein the plurality of age ranges include two or more of: 0 years to 1 year, 1 year to 3 years, 3 years to 8 years, 8 years to 12 years, 12 years to 18 years, or greater than 18 years in the case of adult congenital heart disease.

[0019] In some aspects, the techniques described herein relate to a method, wherein the plurality of age ranges include two or more of: infancy, toddlerhood, school age, preadolescence, adolescence, and adults.

[0020] In some aspects, the techniques described herein relate to a method, wherein determining using the one or more trained models includes determining using one or more trained models were trained with training data including age data and / or sex data from the plurality of prior pediatric patient or adult congenital patients, wherein each of the age data and / or the sex data is combined in the training with the ECG data from the plurality of prior pediatric patient or adult congenital patients.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0021] In some aspects, the techniques described herein relate to a method, wherein determining using the one or more trained models includes determining using one or more trained models were trained with training data including echocardiogram data from the plurality of prior pediatric patient or adult congenital patients, wherein the echocardiogram data is combined with the ECG data from the plurality of prior pediatric patient or adult congenital patients.

[0022] In some aspects, the techniques described herein relate to a method, wherein the echocardiogram data of the training data includes one or more of a death date, ventricular ejection fraction EF, a ventricular mass, a ventricular mass / volume, and a ventricular end-diastolic volume.

[0023] In some aspects, the techniques described herein relate to a method, wherein determining using one or more trained models trained with ECG data from the plurality of prior pediatric patient or adult congenital patients includes determining using one or more trained models trained with filtered ECG data for the plurality of prior pediatric patient or adult congenital patients, the filtered ECG data including noise exceeding a threshold.

[0024] In some aspects, the techniques described herein relate to a method of predicting a cardiac abnormality in a pediatric patient or adult congenital patient, the method including: determining, based on received ECG data from a pediatric patient or adult congenital patient, a prediction of whether the pediatric patient or adult congenital patient has any of one or more cardiac abnormalities, the determining the prediction including analyzing the received ECG data using one or more trained models, wherein the one or more trained models were trained with training data including ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert; and outputting the prediction of whether the pediatric patient or adult congenital patient has any of the one or more cardiac abnormalities.

[0025] In some aspects, the techniques described herein relate to a method of predicting cardiac abnormalities in pediatric patient or adult congenital patients, the method including: determining, based on first received ECG data from a first pediatric patient or adult congenital patient, a first prediction of whether the first pediatric patient or adult congenital patient has one or more cardiac abnormalities, the determining including analyzing the first received ECG data using a trained model; determining, based on second received ECG data from a second pediatric patient or adult congenital patient, a second prediction of whether the second pediatric patient or adult congenital patient has the one or more cardiac abnormalities, the determining including analyzing the second received ECG data using the trained model; and outputting the first prediction and the second prediction for the first pediatric patient or adult congenital patient and the second pediatric patient or adult congenital patient, wherein the trained model was trained with training data including ECGPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 data and cardiac abnormality data from a plurality of prior pediatric patient or adult congenital patients including patients from multiple age ranges of pediatric patient or adult congenital patients, the multiple age ranges including at least one adolescent age range and at least one preadolescent age range, and wherein the first pediatric patient or adult congenital patient has an age in the at least one adolescent age range and the second pediatric patient or adult congenital patient has an age in the at least one preadolescent age range, and wherein the one or more trained models were trained with training data including ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert.

[0026] In some aspects, the techniques described herein relate to a system for predicting a cardiac abnormality in a pediatric patient or adult congenital patient, the system including: an ECG monitoring device; a controller including at least one processor, at least one storage medium having encoded thereon executable instructions that, when executed by the at least one processor, cause the at least one processor to carry out a method including: determining, based on received ECG data from a pediatric patient or adult congenital patient, information regarding whether the pediatric patient or adult congenital patient has one or more cardiac abnormalities, the determining including analyzing the received ECG data using one or more trained model, and wherein the one or more trained models were trained with training data including ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert; and outputting the information regarding whether the pediatric patient or adult congenital patient has the one or more cardiac abnormalities.

[0027] In some aspects, the techniques described herein relate to at least one storage medium having encoded thereon executable instructions that, when executed by at least one processor, cause the at least one processor to carry out a method including: determining, based on received ECG data from a pediatric patient or adult congenital patient, information regarding whether the pediatric patient or adult congenital patient has one or more cardiac abnormalities, the determining including analyzing the received ECG data using one or more trained model, and wherein the one or more trained models were trained with training data including ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert; and outputting the information regarding whether the pediatric patient or adult congenital patient has the one or more cardiac abnormalities.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0028] Methods and systems of the presently disclosed embodiments were isolated or otherwise manufactured in connection with the examples provided below. Other features and advantages will be apparent from the detailed description, and from the claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component that is illustrated in various figures is represented by a like numeral. For purposes of clarity, not every component may be labeled in every drawing. In the drawings::

[0030] FIG. 1 is a block diagram of a system with which some embodiments may operate for analyzing ECG data for a pediatric or adult congenital patient with respect to a cardiac phenotype;

[0031] FIGS. 2A-2B are flowcharts of processes that may be implemented in some embodiments to evaluate information for identifying features related to cardiac phenotypes using ECG data from a pediatric or adult congenital patient, and for training one or more models to predict a cardiac abnormality, respectively;

[0032] FIG. 3 is a block diagram of a computing device with which some embodiments may operate;

[0033] FIG. 4 is a Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) diagram showing initial patient selection, filtering at each data processing stage, patient partitioning, and primary outcome rates. Abbreviations: quality control (QC); congenital heart disease (CHD); left ventricle (LV);

[0034] FIG. 5 shows pediatric electrocardiogram-based deep learning model performance during internal testing. Performance of the artificial intelligence-enhanced pediatric electrocardiogram (AI-pECG; blue) and AI-pECG with age and sex (AI-pECG + age + sex; orange) model performances evaluated during internal testing with receiver operating (left) and precision-recall (right) curves for the: (A) left ventricular (LV) composite; (B) LV dysfunction; (C) LV hypertrophy; and (D) LV dilation outcomes. In panel C, the grey dot represents the benchmark of pediatric electrophysiologist (EP) expert ECG-based diagnosis of LV hypertrophy. AUROC and AUPRC metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping. Abbreviations: positive predictive value (PPV);

[0035] FIG. 6 shows model performance for single random ECGs per patient during internal testing. Performance of the artificial intelligence-enhanced pediatric electrocardiogram (AI-pECG; blue) and AI-pECG with age and sex (AI-pECG + age + sex; orange) model performances evaluated during internal testing on single random ECGs per patient with receiver operating curvesPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 for the: (A) left ventricular (LV) composite; (B) LV dysfunction; (C) LV hypertrophy; and (D) LV dilation outcomes. AUROC metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping;

[0036] FIG. 7 shows pediatric electrocardiogram-based deep learning model performance during testing in the emergency department. Performance of the artificial intelligence-enhanced pediatric electrocardiogram (AI-pECG; blue) and AI-pECG with age and sex (AI-pECG + age + sex; orange) model performances evaluated in a specific clinical setting (emergency department) with receiver operating (left) and precision-recall (right) curves for the: (A) left ventricular (LV) composite; (B) LV dysfunction; (C) LV hypertrophy; and (D) LV dilation outcomes. In panel C, the grey dot represents the benchmark of pediatric electrophysiologist (EP) expert ECG-based diagnosis of LV hypertrophy. AUROC and AUPRC metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping. Abbreviations: positive predictive value (PPV);

[0037] FIG. 8 shows model performance for single random ECGs per patient during testing in the emergency department. Performance of the artificial intelligence-enhanced pediatric electrocardiogram (AI-pECG; blue) and AI-pECG with age and sex (AI-pECG + age + sex; orange) model performances evaluated during testing on single random ECGs per patient with receiver operating curves for the: (A) left ventricular (LV) composite; (B) LV dysfunction; (C) LV hypertrophy; and (D) LV dilation outcomes. AUROC metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping;

[0038] FIG. 9 shows model performance to predict quantitative cutoffs of left ventricular function and remodeling. Performance of the artificial intelligence-enhanced pediatric electrocardiogram (AI-pECG) algorithm evaluated in the internal (left) and emergency department (right) cohorts using receiver operating curves for the following outcomes: (A) left ventricular (LV) composite outcome (LV ejection fraction (LVEF) z-score < -X or LV mass z-score > X or LV end-diastolic volume (LVEDV) z-score > X, where cutoff X = 2.5 (black), 4.0 (blue), 6.0 (red); (B) LV dysfunction (LVEF z-score < -X); (C), LV hypertrophy (LV mass z-score > X); and (D) LV dilation (LVEDV z-score > X). AUROC metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping;

[0039] FIG. 10 shows model performance in age and sex subgroups. Forest plot showing AI- pECG area under the receiver operating curve (AUROC) performance when stratifying by age (age < 1, 1 < age < 3, 3 < age < 8, 8 < age < 12, age > 12) and sex for the following outcomes: left ventricular (LV) composite, LV dysfunction, LV hypertrophy, and LV dilation. 95% confidence intervals are shown using bootstrapping;PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0040] FIG. 11 shows explainability of AI-pECG predictions. Visualization of median waveforms generated in each lead using ECGs from the 100 highest (red) and 100 lowest (green) AI-pECG predictions of left ventricular (LV) dysfunction, hypertrophy, and dilation. Saliency mapping demarcates regions of the ECG waveform having greatest (dark blue) and least (light blue) influence on each outcome. Saliency was averaged over the 100 highest predicted ECGs for each outcome.

[0041] FIG. 12 shows explainability of AI-pECG predictions for the composite outcome. Visualization of median waveforms generated in each lead using ECGs from the 100 highest (red) and 100 lowest (green) AI-pECG predictions of the composite outcome. Saliency mapping demarcates regions of the ECG waveform having greatest (dark blue) and least (light blue) influence on the composite outcome. Saliency was averaged over the 100 highest predicted ECGs for the composite outcome.

[0042] FIG. 13A-C illustrate an AI-ECG model to predict LV and RV dysfunction and dilation in patients with and without Congenital heart disease (CHD), according to exemplary embodiments of this disclosure.

[0043] FIG. 14 illustrates a STROBE diagram showing initial patient selection and the final cohort.

[0044] FIGS. 15A-D illustrate external validation of Electrocardiogram-Based Deep Learning Model Performance.

[0045] FIG. 16 illustrates ECG-based deep learning model performance on random ECG- Cardiac magnetic resonance (CMR) pairs.

[0046] FIG. 17 illustrates ECG-based deep learning model performance for detecting Biventricular Dysfunction.

[0047] FIG. 18 illustrates Al -ECG subgroup model performance when stratifying by age and sex.

[0048] FIG. 19 illustrates sensitivity analysis of overall model performance excluding each lesion.

[0049] FIG. 20 illustrates performance of ECG-based deep learning model in Tetralogy of Fallot.

[0050] FIG. 21 illustrates sensitivity analysis of right ventricle (RV) at-risk model performance to excluding each RV at-risk lesion.

[0051] FIG. 22 illustrates performance of RV dilation-specific model in Tetralogy of Fallot.

[0052] FIG. 23 survival analysis in Tetralogy of Fallot Based on AI-ECG Predictions.

[0053] FIGS. 24A-C illustrate Survival Analysis Based on AI-ECG Predictions.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0054] FIG. 25 illustrates an explainability of Artificial Intelligence-EnhancedElectrocardiogram Outcome Predictions.

[0055] FIG. 26 illustrates internal testing and external validation of the AI-ECG model to predict left ventricular systolic dysfunction.

[0056] FIGS. 27A-B illustrate model performance across congenital heart disease lesion subgroups.

[0057] FIGS. 28A-B illustrate future left ventricular systolic dysfunction or mortality based on AI-ECG classification. (FIG. 28A) Incidence of future LVEF <40% for the overall test cohort (left), patients with cardiomyopathy (middle), and patients with tetralogy of Fallot (right) initially with LVEF >40%, stratified by initial network classification (low-risk in green, intermediate-risk in blue, and high-risk in red). (FIG. 28B) Survival analysis when stratifying patients as low-risk, intermediate- risk, or high-risk based on AI-ECG left ventricular systolic dysfunction probabilities for the overall cohort (left), cohort with cardiomyopathy (middle), or cohort with tetralogy of Fallot (right).

[0058] FIG. 29 illustrates saliency mapping and median waveform analysis to generate hypotheses of ECG features driving LVSD model predictions.

[0059] FIGS 30A-30B illustrates AI-ECG model performance to detect ECG abnormalities. FIG. 30A illustrates performance of the artificial intelligence-enhanced ECG model alone and with demographic data evaluated in the testing cohorts using receiver operating (AUROC) and precision recall (AUPRC) curves for the following outcomes: any abnormality, Wolff Parkinson White syndrome (WPW), and prolonged QTc. FIG. 30B illustrates incidence of future abnormal ECG, WPW, or prolonged QTc for the test cohort stratified by initial network classification.

[0060] FIG. 31 illustrates model performance across age and sex subgroups. Forest plots showing area under the receiver operating (AUROC; red) and precision recall (AUPRC; black) curve performance when stratifying by age, sex, and reader for diagnosing any abnormality (left), Wolff Parkinson White syndrome (WPW; middle) and prolonged QTc (right).

[0061] FIGS. 32A-32B illustrate model classification readjudication results of ECGs deemed as false positives (AI-ECG positive but baseline reader negative; FP) or false negatives (AI-ECG negative but baseline reader positive; FN). FIG. 32A illustrates a heatmap of classifications (positive = dark; negative = light) by baseline reader (Baseline), AI-ECG, and four expert readers. Hierarchical clustering of AI-ECG and readers shown above. Fleiss K (to assess agreement across experts) shown below. FIG. 32B illustrates Cohen’s K for agreement between the baseline reader and expert readjudicator (Baseline-Expert Agreement; black) versus the AI- ECG and expert readjudicator (AI-ECG-Expert Agreement).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0062] FIG. 33 illustrates model explainability through visualization of high-risk (red) and low- risk (green) median waveforms for WPW (top) and prolonged QTc (bottom).

[0063] FIGS. 34A-C illustrate results from using Infant AI-ECG to predict critical congenital heart disease. FIG. 34A illustrates performance of the infant AIO-ECG model evaluated during internal (blue) and external (orange) testing with receiver operating (AUROC; left) and precision-recall (AUPRC; right) curves. FIG. 34B illustrates lesion-specific AUROC and AUPRC performance during internal (blue) and external (orange) testing. Color-coded prevalence of each lesion inset below. FIG. 34C illustrates internal testing of lesion-specific AUROC (black) and AUPRC (red) performance. Prevalence of each lesion inset below.

[0064] FIG. 35 illustrates results for neonatal model explainability. Visualization of neonatal (< 7 days old) lesion-specific median waveforms generated in each lead using ECGs from the 25 highest (red) and 25 lowest (green) AI-ECG predictions. Saliency mapping demarcates regions of the ECG waveform having greatest (dark blue) and least (light blue) influence on predicting each lesion. Abbreviations: Complete atrioventricular canal defect (CAVC); coarctation of the aorta (CoA); pulmonary atresia (PA); tetralogy of Fallot (ToF); hypoplastic left heart syndrome (HLHS).

[0065] FIGS. 36A-B illustrates performance in the (left) overall and (right) paired ECG-CMR test cohorts for predicting the 5-year risk of SCD. FIG. 36A illustrates AUROC of AI-ECG to predict the 5-year risk of SCD. Model benchmarked to (left) QRS interval duration and (right) BVGFI (right). FIG. 36B illustrates (left) freedom from SCD Kaplan-Meier analysis when stratifying patients using AI-ECG only, and (right) freedom from SCD Kaplan-Meier analysis when stratifying patients based on risk factors.

[0066] FIG. 37 illustrates internal testing and external validation study populations comprising BCH patients enrolled in the INDICATOR (International Multicenter TOF Registry), according to exemplary embodiments of this disclosure.

[0067] FIG. 38A-B illustrate (FIG. 38A) Artificial intelligence-enhanced electrocardiogram (AI-ECG) model performance to predict 5-year mortality' during internal testing (blue) and external validation (orange), and (FIG. 38B) Kaplan-Meier curve survival analysis of internal (left) and external (right) cohorts when stratifying patients as low- (green) or high-risk (red) based on AI-ECG predictions.

[0068] FIG. 39A-B illustrate (FIG. 39 A) benchmarking of AI-ECG performance in predicting the primary outcome to established rTOF risk stratification markers; and (FIG. 39B) survival assessment after ECG-CMR pair by stratifying patients into low- or high-risk AI-ECG groups .

[0069] FIG. 40 illustrates internal testing (left) and external validation (right) AI-ECG probabilities of 5-year mortality' (y-axis) as a function of age at ECG (x-axis).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0070] FIG. 41 illustrates visualization of median waveforms generated in each lead using ECGs from the 50 highest (red) and 50 lowest (green) AI-ECG predictions of the internal (top) and external (bottom) cohorts.

[0071] FIGS. 41A-B illustrate saliency mapping and median waveform analysis to interpret model behavior.

[0072] FIG. 42 illustrates an AI-Enhanced ECG to predict mortality in repaired Tetralogy of Fallot.

[0073] FIGS. 43A-C illustrate Kaplan-Meier survival analysis of training and testing cohorts.

[0074] FIGS. 44A-C illustrate ECG-based deep learning model performance.

[0075] FIG. 45 illustrates model performance in congenital heart disease subgroups.

[0076] FIG. 46 illustrates Kaplan-Meier survival analysis based on artificial intelligence- enhanced electrocardiogram risk stratification.

[0077] FIG. 47 illustrates Lesion-specific Kaplan-Meier survival analysis based on artificial intelligence-enhanced electrocardiogram risk stratification.

[0078] FIG. 48 illustrates explainability of artificial intelligence-enhanced electrocardiogram predictions.

[0079] FIG. 49 provides a central illustration of a large and diverse pediatric and adult congenital heart disease cohort used to train and test an artificial intelligence-enhanced electrocardiogram algorithm to accurately predict 5-year mortality across a range of congenital heart disease lesions.

[0080] While the above-identified drawings set forth presently disclosed embodiments, other embodiments are also contemplated, as noted in the discussion. This disclosure presents illustrative embodiments by way of representation and not limitation. Numerous other modifications and embodiments can be devised by those skilled in the art which fall within the scope and spirit of the principles of the presently disclosed embodiments.DETAILED DESCRIPTION

[0081] Disclosed herein are techniques for evaluating cardiac health of a pediatric or adult congenital patient, including techniques for generating one or more models (e.g., machine learning models) trained to predict a cardiac abnormality in a pediatric or adult congenital patient. Some methods may include receiving electrocardiogram (ECG) data of a pediatric or adult congenital patient, determining whether the patient corresponds to one or more of multiple cardiac phenotypes, each corresponding to a presence of a cardiac abnormality based on analysis by one or more trained models, and outputting the cardiac phenotype(s) based on the ECG data. In some embodiments, the cardiac abnormality may include one or more of mortality, ventricularPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 dysfunction, ventricular dilation, ventricular hypertrophy, one or more rhythm abnormality such as Wolff Parkinson White syndrome or a prolonged QTc interval, or critical congenital heart disease. In embodiments that use one or more trained models, the trained model(s) may be trained using training data that includes ECG data from multiple prior pediatric or adult congenital patients and information indicating whether each prior patient had one or more cardiac abnormalities.

[0082] ECG is a quantitative method that indicates pulse waveforms in cardiac screening. While ECG has been used for some forms of cardiac screening, ECG has not been reliable for cardiac phenotyping or diagnostics of cardiac abnormalities based on a specific correlation between waveforms and cardiac phenotypes, due to variability in waveforms. As such, for cardiac phenotyping to diagnose a cardiac abnormality, a patient has conventionally relied on MRI or on echocardiograms instead. These techniques have high reliability and have been in wide use, and clinicians therefore prefer to use these techniques for indications including: known acquired / congenital heart disease, suspicion for arrhythmia or structural heart disease, sports clearance, medication initiation / monitoring, and others. The growing volume of ECGs, along with considerations for universal ECG screening, underscores the need for rapid and reliable ECG interpretations. Unfortunately, availability of this equipment and diagnostic technique can be limited and have a high cost, particularly in smaller clinic settings or remote areas. This means echocardiogram or MRI may not be available for some patients at all (where the technology is not present) or may not be available to certain patients (e g., where limited availability imposes cost or other hurdles to use). The inventors have recognized that there would be advantages to techniques for cardiac phenotyping and diagnostics using other techniques.

[0083] Some recent work has been performed in experimenting with deep learning-based artificial intelligence-enhanced ECG (AI-ECG) algorithms for use in diagnostics for certain cardiac abnormalities. AI-ECG studies have been explored for use in adult populations, such as to predict a range of general adult cardiovascular abnormalities. However, existing methods show variability in diagnostic efficacy, raising challenges for use. Further, previous studies have primarily focused on adults with acquired cardiovascular diseases and have not been directed to adults with congenital heart disease, generally due to a number of reasons including, for example, insufficient data for testing and a lack of likelihood that their methods would be similarly applicable. Regarding the latter, in particular, there are ECG signatures in adult congenital heart disease that are heterogeneous and completely distinct from the general adult population (e.g., adults with acquired cardiovascular diseases), making it nonobvious that an AI-ECG model would function successfully in this population. Therefore, it is reasonable to expect that existing AI-ECG models for the general adult population would not translate well for the adult congenital heart disease populations.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0084] While there has been some early work on use of ECG in adult cardiovascular diagnostics, it has not been explored for pediatric or adult congenital use, including based on significant anatomical differences between adult and pediatric or adult congenital patient populations and within pediatric or adult congenital patient populations. These significant anatomical differences have suggested that even if ECG were ultimately be found to be a viable diagnostic tool for adult populations, ECG would not necessarily be successful as a diagnostic for pediatric and adult congenital populations. For example, the epidemiology and patterns of normal versus abnormal pediatric and adult congenital ECG waveforms differ significantly from those of structurally normal adult heart, which prevents applicability of adult AI-ECG algorithms to pediatric and adult congenital cohorts. Further, there are well-known progressive anatomical and physiological changes occurring from birth to adolescence leading to more age-dependent variations in pediatric ECGs than in adult ECGs, preventing applicability of observable changes in one age population to other age populations even within a pediatric group.

[0085] ECG data is not traditionally used to deeply phenotype the cardiac health of pediatric and adult congenital population, thus any implementation of Al-enhanced algorithms would be expected to exhibit similar unreliability due to the anatomical data on which such algorithms would rely for training. This prevents the ability to understand ECG data across multiple age ranges, in particular those in the pediatric age range, to provide guidance for how to selectively identify whether a pediatric or adult congenital patient may be exhibiting a cardiac phenotype, such as those mentioned above, that are indicative of a cardiac abnormality. Even if one were to apply an AI- ECG model to pediatric and adult congenital cohorts, the variability across age ranges such as infancy and adolescence would prevent reliability of such a model to any individual within pediatric cohorts, i.e., from 0 to 18 years of age. The presence of congenital heart disease provides an additional underlying hurdle of unreliability, as the abnormalities in the heart of a pediatric or adult congenital patient - which may vary from mild to severe - results in even further variability in the resulting ECG waveforms. As a result, there are relatively few available AI-ECG applications to pediatric and adult congenital cardiology, highlighting the paucity of pediatric AI- ECG models to date that could benefit both resource-rich and resource limited pediatric and adult congenital settings.

[0086] The inventors have recognized and appreciated, however, that an ECG-based analysis for pediatric or adult congenital patients would be advantageous, including in healthcare scenarios where MRI or echocardiogram equipment may be unavailable or difficult to obtain or use. The inventors determined, for example, that training a machine learning model for use with pediatric and adult congenital ECG data (sometimes termed herein an “AI-pECG model”) with ECG data paired with echocardiogram data and / or MRI data from pediatric populations may be used toPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 predict cardiac phenotypes for pediatric or adult congenital patients. Such cardiac phenotypes may include risk of mortality, LV dysfunction, hypertrophy, and dilation. Echocardiograms are highly useful for diagnosing various cardiac abnormalities, but are associated with high costs and may be referred only when necessary. Further, while previous diagnostic tools have been useful for identifying particular cardiac abnormalities in isolation, there are no tools available for pediatrics in determining such abnormalities either individually or as a composite outcome. In this context, a composite outcome would include a prediction of at least two of the aforementioned phenotypes.

[0087] The inventors have further recognized and appreciated the value of obtaining dedicated and age-delineated data sets that allow training for each respective age group within pediatric populations. In particular, for one experiment, the inventors accumulated nearly 100,000 ECG- echo pairs < 2 days apart, an independent internal test set of >20,000 ECG-echo pairs, as well as in a test set of patients in the emergency room with >3,000 ECG-echo pairs. Subgroup analysis was also performed in this experiment, using age partitioning with groupings of age < 1, 1 < age < 3, 3 < age < 8, 8 < age < 12, and 12 < age < 18 years, matching groupings related to infancy, toddlerhood, school age, preadolescence, and adolescence.

[0088] The inventors recognized and appreciated the value of filtering training data to eliminate some cardiac phenotypes. For example, some embodiments may include data filtering that excludes data (e.g., ECG waveforms with poor qualify data; patients with major congenital heart disease based on their electronic health record) from pediatric patients with major congenital heart disease, thus ensuring models are trained without information for patients with major congenital heart disease. This may prevent a model from predicting presence of major congenital heart disease, but may have reliability advantages for predicting other cardiac phenotypes.

[0089] Human expert interpretation of echocardiograms of LV function and remodeling is particularly useful for training an AI-pECG model, given -25% of the accumulated patient data are without quantitative measures (substantially reducing available training datasets for the AI- pECG), and a human expert opinion incorporates multiple clinical datapoints that aid in decision making for use in training the model. However, because qualitative and quantitative cutoffs individually have their own drawbacks, and most emergency medicine physicians will attempt cardiac point-of-care ultrasound and report qualitative outcomes on function, the inventors wanted to develop a model that worked regardless of a cutoff method. Some embodiments described herein may aid in obtaining data with greater accuracy and qualify for training a machine learning model without such cutoffs, which may increase a reliability thereof.

[0090] The inventors have further recognized and appreciated the clinical utility and convenience of AI-pECG modeling. Specifically, some models described herein may be performed with cheap and rapidly generated ECG waveform data. The robust performance of models suggests they arePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 at least partially resistant to noise generated from obtaining data. Importantly, in some embodiments, each model may implement only a single modality (i.e., independent of demographics such as age / sex), rather than a complex clinical scoring system that would require user interaction and may be susceptible to input error.

[0091] It should further be noted there is a relative lack of scoring systems for ventricular dilation or dysfunction in pediatric or adult congenital patients, thus enhancing the need for such a model. The inventors have further identified through saliency mapping as to regions of pediatric ECG waveforms that influence model predictions and provide novel insight into clinicians detecting risk of mortality or LV dysfunction and remodeling. In particular, the use of saliency mapping in some implementations may allow for visual comparison of ECGs with existing algorithms for ECG interpretation, with outcomes, both for individual and composite phenotype predictions, being representing of an underlying outcome. The inventors have appreciated, for example, that to predict LV hypertrophy, salient features in some cases may be or include precordial and limb lead I QRS complexes, with deep S waves in V1-V2 and high amplitude R waves in limb lead I included as predictive high-risk features. Further, 15or LV dilation, the salient features in some cases may be or include lateral precordial (V4-V6) QRS complexes, with high amplitude R waves in V4-V6 included as predictive high-risk features. Saliency mapping for the composite outcome was observed to merge features from each of the individual outcomes of interest, i.e., LV hypertrophy and LV dilation.

[0092] The inventors have further recognized and appreciated the utility of an AI-pECG in an emergency room, expanding additional clinical settings that would benefit from such tools. The economic burden of cardiac abnormalities, when addressed in emergency rooms, includes the potential for misdiagnosis leading to unnecessary referrals and associated costs of echocardiograms, as well as missed potentially fatal diagnoses. AI-pECG predictions could help guide an emergency physician’s need to consult a pediatric cardiologist, and could help guide a pediatric cardiologist’s decision of if and / or when to get an echocardiogram for a child without congenital heart disease. This democratization of specialty expertise is likely to be particularly valuable for hospitals with low pediatric volumes and / or limited pediatric cardiology experience.

[0093] The inventors have further recognized and appreciated the variance of expertise in interpreting ECGs, thus requiring more reliability in the training data used for training Al-based ECG (AI-ECG) models. It is generally known that human expertise is required for pediatric ECG interpretation to account for the progressive anatomic / physiologic changes from the newborn stage to adulthood, which corresponds to significantly different ECG patterns and epidemiology. While ECG vendors have implemented rule-based algorithms, these are generally focused on adult literature with limited standalone value in the pediatric ECG setting. There is value,PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 especially in low-resource settings with limited pediatric cardiology expertise, to have access to automated ECG diagnostic tools designed for pediatric use, but with improve reliability over existing models. There remains a paucity of available AI-ECG applications to pediatric and congenital cardiology given: 1) the lack of big data hindering similar research; and that 2) AI- ECG algorithms for adults are expected to have poor generalizability to pediatric cohorts. Thus, there are currently no AI-ECG algorithm for pediatric ECG diagnosis with sufficient reliability. The inventors have identified this gap by training and testing an AI-ECG model using >500,000 ECGs on >200,000 patients to reliably identify common and rare ECG findings in the pediatric population. By benchmarking model performance to commercial software and readjudicating misclassified tracings of initial ECG reads to expert pediatric electrophysiologists, the inventors have strategically incorporated additional data-based safeguards to overcome gaps in AI-ECG model performance.

[0094] Techniques described herein may be useful in some embodiments in generating a model to output one or more cardiac phenotypes determined for a pediatric or adult congenital patient based on received ECG data. Further described herein are examples of techniques and systems with which such techniques may be used. These include, for example, (1) systems with which some embodiments of the methods described herein may operate; (2) methods for conducting training of at least one model using information regarding one or more cardiac phenotypes corresponding to a presence of a cardiac abnormality; (3) methods of identifying one or more cardiac phenotypes predicted by the at least one model; and (4) methods for training a model using additional data or information from the pediatric or adult congenital patient, such as echocardiogram data, age range, and / or sex, to detect a cardiac abnormality or a composite of multiple cardiac abnormalities predicted by the model to be indicative of one or more cardiac abnormalities.

[0095] The following description and examples illustrate in detail some embodiments of techniques and technologies described herein. It is to be understood that embodiments are not limited to acting in accordance with the specific examples provided herein, as other approaches are possible. Those of skill in the art will recognize that there may be variations and modifications from the specific examples below that are within the scope of this disclosure.Illustrative Systems

[0096] FIG. 1 illustrates a block diagram of a system 100 with which some embodiments may operate. The system 100 can evaluate a patient(s) to determine whether his or her respective phenotype may be normal / healthy or whether the phenotype corresponds to a risk of a particular cardiac abnormality, such as future mortality, ventricular dysfunction, ventricular dilation, ventricular hypertrophy, one or more rhythm abnormality such as Wolff Parkinson WhitePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 syndrome or a prolonged QTc interval, or critical congenital heart disease. The system 100 may in some embodiments produce the estimate of risk by analyzing a combination of pECG data with a trained model, where the trained model may be trained on prior pECG data. Such prior pECG data may be or include those from previous pediatric or adult congenital patients.

[0097] The system 100 can include a patient 102. In some embodiments, the patient 102 may be healthy, or has symptoms of a cardiac phenotype corresponding to the presence of a cardiac abnormality such that testing may be done to determine whether the patient 102 has the cardiac abnormality. For example, the patient 102 can be a recipient of health care services that are administered by healthcare professionals. For example, the patient 102 can be ill or injured and require treatment. The patient 102 can seek the advice of healthcare professionals regarding treatment, which may, for example, be a response to pain or a feeling of unease. Accordingly, one or more clinicians 104 may interface with the patient 102 to manage illness or injury of the patient. Examples of clinicians 104 include a physician, nurse, physician assistant, nurse practitioner, psychologist, clinical pharmacist, clinical scientist, or specialist physician such as an electrophysiologist.

[0098] One or more ECG recordings 106 may be obtained from the patient 102, such as by the clinician 104 obtaining the sample through an ECG monitor 107 from the patient 102, though it should be appreciated that embodiments are not so limited. In some embodiments, the ECG recording 106 may be sufficient to produce signal data for the patient 102. More generally, the ECG recording 106 can be gathered from the patient to aid in medical diagnostics or evaluation of treatment. The ECG recording 106 can be obtained prior to use, which can include receiving from a database.

[0099] The system 100 can include a pECG data analysis facility 116, which may be or include one or more tools for analyzing the ECG recording 106. The exact form of the pECG data analysis facility 116 may depend on the form of the ECG recording 106 taken from the patient. Examples of analytical tools for detecting markers were discussed above, any one, two, or more of which may be implemented in some embodiments. The pECG data analysis facility 116 may also include a signal analyzer and / or an ECG analysis module.

[0100] The system 100 can include a client computing device 110, which may be a desktop or laptop personal computer, smart mobile phone, server, or other suitable device. The client computing device 110 may include a client interface 112 by which the patient 102 or the clinician 104 may interact with the client computing device 110. For example, the patient 102 orthe clinician 104 can use the client interface 112 to interface with the pECG data analysis facility 116 of the server computing device 114. For example, the patient 102 and / or clinician 104 may operate the client interface 112 to initiate analysis of the ECG recording 106 by the pECG data analysis facilityPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 116 and display analysis results such as whether features were detected and / or levels of those features in the interface 112. The patient 102 and / or clinician 104 may additionally or alternatively operate the client interface 112 to input features and / or feature levels obtained from the pECG data analysis facility 116, such as output to the patient 102 and / or clinician 104 in another interface. Those values may be provided to the pECG data analysis facility 116. As a further example, the patient 102 and / or clinician 104 may operate the client interface 112 to initiate analysis of the ECG recording 106 by the pECG data analysis facility 116. Results of analysis of the results (received from the interface 112) by the pECG data analysis facility 116 may be output to the client interface 112, such as by being received at the client interface 112 and displayed on the device 110. In some embodiments, as mentioned above, the client interface 112 may include a web interface, such as one or more web pages into which values may be output and which may display results of the analysis by the pECG data analysis facility 116, but embodiments are not so limited. The client interface 112 may accept input in a variety of different formats, such as through speech recognition, text input, or other means, as embodiments are not limited in this respect.

[0101] The system 100 can include a server computing device 114, which may include a pECG data analysis facility 116 configured to analyze factors (e.g., derived from the ECG recording 106) for the patient 102 with one or more cardiac phenotype to determine a risk that the patient 102 has a cardiac abnormality. These factors may include demographic information such as age or sex of patient 102 and may include information represented in a numeric form or as change indicators, such as trajectory indications.

[0102] The system 100 can include a network 118 to facilitate communications among the pECG data analysis facility 116, the client computing device 110, and the server computing device 114. The network 118 can be or include any one or more wired and / or wireless, local- and / or wide- area network, including one or more enterprise networks and / or the Internet.

[0103] While the example of FIG. 1 includes the client interface on a device 110 separate from the pECG data analysis facility 11 , it should be appreciated that embodiments are not so limited. In other embodiments, the client interface 112 may be an interface of the pECG data analysis facility 116 and may be operated by the patient 102 and / or the clinician 104. Additionally or alternatively, while the pECG data analysis facility 116 is illustrated on a different computing device from the client computing device 110, embodiments are not so limited. In some embodiments, the client interface 112 may not be separate from the pECG data analysis facility 116, but instead may be implemented as a single program or software application. In some embodiments, a pECG data analysis facility 116 may include the client interface 112, and the interface 112 and facility 116 may be implemented within the same program or application executed on the pECG data analysis facility 116.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025Example Methods of predicting a cardiac abnormality

[0104] FIGS. 2A-2B illustrate flowcharts of processes 1000 and 2000 that may be implemented in some embodiments to predict a cardiac abnormality and train a model for predicting a cardiac abnormality, respectively. Processes 1000 and 2000 can be implemented in some embodiments by the pECG data analysis facility 116 of the server computing device 114, which can output one or more cardiac phenotypes that satisfy predetermined criteria for the ECG data.

[0105] In step 1001, the pECG data analysis facility 116 receives ECG data from a pediatric or adult congenital patient, referred to herein as pECG data. The pECG data may include one or more of ECG recordings 106 and stored ECG data on a computer-readable storage medium. In some embodiments, the pECG data includes raw ECG signals exported from an ECG data management system. For example, the raw ECG signals may include waveform data, and the waveform data may include a vector of data sampled at a rate of 250 Hz for 10 seconds duration (2500 samples) corresponding to a lead (I, II, and V1-V6). The vector of data may be one-dimensional. The vector of data may be linearly transformed. In some embodiments, the transformation is based on the Einthoven law and / or Goldberger equation, and may be used to obtain leads III, aVF (augmented vector foot), aVL (augmented vector left), and aVR (augmented vector right). For example, the waveform data includes voltages that have been one or more of filtered and digitized for sampling and linear transformation. Given that ECGs are prone to recording errors (e.g., baseline wander; electrical interference), a high pass filter may be utilized. In some embodiments, the high pass filter may include a cutoff frequency of about 0.8 Hz, a rejection band of about 0.2 Hz, a ripple in a passband of about 0.5 dB, and an attenuation in a rejection band of about 40 dB. In some embodiments, the ECG may be trimmed to facilitate conveniently working with convolution neural networks. For example, the ECG data may be trimmed to about 2048 samples (approximately 8 seconds). In addition to analyzing pECG data for a patient, the pECG data analysis facility 116 may analyze demographic data for corresponding patients. Such demographic data may include age of the patient, sex of the patient, socioeconomic information regarding the patient, or other demographic information. The pECG data may include measurements including one or more of QRS interval, QRS axis, T axis, P axis, PR interval, QT interval, QTc, and heart rate.

[0106] In step 1002, the pECG data analysis facility 116 analyzes the pECG data (and, in some embodiments, demographic data or other patient data) using at least one trained model disclosed herein, wherein the at least one trained model may have been trained according to process 2000 described in more detail below. The pECG data may be sent to the same trained model,PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 regardless of the patient’s age or sex, as the model may be used for a variety of age ranges for pediatric or adult congenital patients. Based on the analysis in step 1002, one or more cardiac phenotypes may be determined, such that step 1003 includes assigning one or more cardiac phenotypes to the pediatric or adult congenital patient. In some embodiments, the one or more phenotypes include a normal (i.e., healthy) phenotype, a ventricular dysfunction phenotype, a ventricular dilation phenotype, a ventricular hypertrophy phenotype, a risk of mortality phenotype, one or more rhythm abnormality such as Wolff Parkinson White syndrome or a prolonged QTc interval, and critical congenital heart disease. The one or more cardiac phenotypes are preprogrammed to correspond to a presence of a cardiac abnormality. The determining may include determining a probability of the pediatric or adult congenital patient having any of one or more cardiac abnormalities, or each of the one or more cardiac abnormalities. The determining may further include determining a probability of the pediatric or adult congenital patient having the one or more cardiac abnormalities based on the one or more cardiac phenotypes. The determining may also include assigning a qualitative assessment for each of the one or more cardiac phenotypes, such that the qualitative assessment indicates whether a severity of a cardiac abnormality is indicated by the cardiac phenotype.

[0107] In some embodiments, the probability of the pediatric or adult congenital patient having the one or more cardiac abnormalities may further be used to predict a mortality of the pediatric or adult congenital patient, such that the mortality is associated with one or more of the presence of the one or more cardiac abnormality and the probability of the pediatric or adult congenital patient having the one or more cardiac abnormality'. In some embodiments, the model may be trained to directly predict the risk of mortality of the pediatric or adult congenital patient.

[0108] The determining may include determining a probability of the pediatric or adult congenital patient having each of the one or more cardiac abnormalities, such that each cardiac abnormality may individually be associated with its own probability. The cardiac abnormalities may include one or more of ventricular dysfunction, ventricular dilation, ventricular hypertrophy, one or more rhythm abnormality such as Wolff Parkinson White syndrome or a prolonged QTc interval, and critical congenital heart disease, where each may be associated with either a left or right ventricle. Determining the one or more cardiac phenotype may include comparing the probability of the patient having each of the one or more cardiac abnormalities to a threshold. The threshold for each of the one or more cardiac abnormality may be determined by the at least one trained model. Finally, step 1003 includes outputting the one or more cardiac phenotypes.

[0109] Process 2000 may be used to generate the at least one trained model used in step 1002 of process 1000. In step 2001, the pECG data analysis facility 116 receives training data from one or more prior pediatric or adult congenital patients. Different training data may be used forPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 training of the one or more trained models. For example, the training data may include ECG data, i.e., training pECG data. The training pECG data may further include ECG data from prior pediatric or adult patients with congenital heart disease. In some embodiments, the training pECG data may be filtered to exclude ECG data from prior pediatric patients with congenital heart disease. In some embodiments, the training pECG data may be ECG data that is filtered of noise exceeding one or more thresholds.

[0110] The training data may also include echocardiogram data associated with the ECG data to form ECG-echo pairs. The echocardiogram data may include one or more of a ventricular ejection fraction EF, a ventricular mass, a ventricular mass / volume, and a ventricular end-diastolic volume. The training data may also include MRI data associated with the ECG data to form ECG- MRI pairs. The prior pediatric or adult congenital patients may be stratified across one or more age ranges, such that each of the one or more age ranges are between 0 and 18 years in being consistent with a pediatric cohort, or greater than 18 years for the adult congenital cohort. For example, the one or more age ranges may include two or more of: 0 years to 1 year, 1 year to 3 years, 3 years to 8 years, 8 years to 12 years, or 12 years to 18 years. In some embodiments, the one or more age ranges includes two or more of: infancy, toddlerhood, school age, preadolescence, adolescence, and adults in the case of adult congenital patients. The one or more prior pediatric or adult congenital patients may further be stratified by their sex.

[0111] In step 2002, the pECG data analysis facility 116 receives information as to whether each of the prior pediatric or adult congenital patients had one or more of the cardiac abnormalities. The prior pediatric or adult congenital patients may include a probability of each of one or more cardiac phenotypes associated with one or more cardiac abnormality. The prior pediatric or adult congenital patients may also include a probability of a composite of each of the one or more cardiac phenotypes.

[0112] In step 2003, the pECG data analysis facility 116 trains at least one model to determine at least one cardiac phenotype using the training data and the information. In some embodiments, the training may incorporate one or more of the sex and the age of each of the prior pediatric or adult congenital patients. Further, each of the age data and / or the sex data is combined in the training with the pECG data from the prior pediatric or adult congenital patients. The training may also combine the echocardiogram data with the pECG data from the prior pediatric or adult congenital patients.

[0113] In some embodiments, process 2000 may further include evaluating a performance of the at least one trained model, wherein the evaluating may include comparing outputs of the at least one trained model to expert diagnoses of each of the prior pediatric or adult congenital patients for the one or more cardiac abnormalities. The evaluating may further include comparing aPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 sensitivity of the at least one trained model to the expert diagnoses. Each of the expert diagnoses may include one or more qualitative and / or quantitative cutoffs for establishing a diagnosis.Some Machine Learning Model Implementations

[0114] As discussed above, the pECG analysis facility 116 may employ various machine learning techniques to identify features in the ECG recordings 106 regarding cardiac phenotypes. These machine learning techniques may, for example, leverage models of the features to be identified in the information. These models may be constructed using various supervised and / or unsupervised learning techniques. In supervised learning, the pECG analysis facility 116 may train a model using training data including the output features to be identified (e.g., LV dysfunction, LV hypertrophy, LV dilation). For example, an supervised learning model includes receiving training data, receiving annotations in the received training data, and training the model using the annotated training data.

[0115] The training data may include, for example, training data that includes the feature(s) that the model will be trained to recognize. For example, the pECG analysis facility may be constructing a model to be used for recognizing LV dysfunction or remodeling on echocardiogram. Similar training can be performed to recognize LV dysfunction or remodeling on MRI.

[0116] The pECG analysis facility 116 may train the model using the annotated training data. In some embodiments, the pECG analysis facility 116 may train the model to identify the features marked in the annotated training data. For example, the training data may include LV dysfunction or remodeling on echocardiogram. In this example, the pECG analysis facility 116 may be trained to identify suspected features indicative of cardiac phenoty pes using the annotated training data.

[0117] Any of a variety of machine learning algorithms may be used in implementing the disclosed image processing and analysis methods. For example, the machine learning algorithm employed may include a supervised learning algorithm, an unsupervised learning algorithm, a semi-supervised learning algorithm, a deep learning algorithm, or any combination thereof. In some embodiments, the machine learning algorithm employed may include an artificial neural network algorithm, a Gaussian process regression algorithm, a logistical model tree algorithm, a random forest algorithm, a fuzzy classifier algorithm, a decision tree algorithm, a hierarchical clustering algorithm, a k-means algorithm, a fuzzy clustering algorithm, a deep Boltzmann machine learning algorithm, a deep convolutional neural network algorithm, a deep recurrent neural network, or any combination thereof, some of which will be described in more detail below.

[0118] As noted above, the machine learning algorithm(s) employed in the disclosed methods and systems for characterizing ECG recordings 106 may include a supervised learningPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 algorithm, an unsupervised learning algorithm, a semi-supervised learning algorithm, a deep learning algorithm, etc., or any combination thereof.

[0119] Having described example processes, it should be appreciated that various alternations may be made to the described processes without departing from the scope of the present disclosure. For example, various acts may be omitted, combined, repeated, or added. Further, the acts in any of the processes described herein do not need to be performed in the particular order shown.Illustrative computer implementations

[0120] Techniques operating according to the principles described herein may be implemented in any suitable manner. Included in the discussion above are a series of flow charts showing the steps and acts of various processes that analyze marker data, such as ECG features and other data, for one or more patients expressing cardiac phenotypes to determine features that may be indicative of a cardiac abnormality. The processing and decision blocks of the flow charts above represent steps and acts that may be included in algorithms that carry out these various processes. Algorithms derived from these processes may be implemented as software integrated with and directing the operation of one or more single- or multi-purpose processors, may be implemented as functionally-equivalent circuits such as a Digital Signal Processing (DSP) circuit or an Application-Specific Integrated Circuit (ASIC), or may be implemented in any other suitable manner. It should be appreciated that the flow charts included herein do not depict the syntax or operation of any particular circuit or of any particular programming language or type of programming language. Rather, the flow charts illustrate the functional information one skilled in the art may use to fabricate circuits or to implement computer software algorithms to perform the processing of a particular apparatus carrying out the types of techniques described herein. It should also be appreciated that, unless otherwise indicated herein, the particular sequence of steps and / or acts described in each flow chart is merely illustrative of the algorithms that may be implemented and can be varied in implementations and embodiments of the principles described herein.

[0121] Accordingly, in some embodiments, the techniques described herein may be embodied in computer-executable instructions implemented as software, including as application software, system software, firmware, middleware, embedded code, or any other suitable type of computer code. Such computer-executable instructions may be written using any of a number of suitable programming languages and / or programming or scripting tools, and also may be compiled as executable machine language code or intermediate code that is executed on a framework or virtual machine.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0122] When techniques described herein are embodied as computer-executable instructions, these computer-executable instructions may be implemented in any suitable manner, including as a number of functional facilities, each providing one or more operations to complete execution of algorithms operating according to these techniques. A “functional facility,” however instantiated, is a structural component of a computer system that, when integrated with and executed by one or more computers, causes the one or more computers to perform a specific operational role. A functional facility may be a portion of or an entire software element. For example, a functional facility may be implemented as a function of a process, or as a discrete process, or as any other suitable unit of processing. If techniques described herein are implemented as multiple functional facilities, each functional facility may be implemented in its own way; all need not be implemented the same way. Additionally, these functional facilities may be executed in parallel and / or serially, as appropriate, and may pass information between one another using a shared memory on the computer(s) on which they are executing, using a message passing protocol, or in any other suitable way.

[0123] Generally, functional facilities include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data ty pes. Ty pically, the functionality of the functional facilities may be combined or distributed as desired in the systems in which they operate. In some implementations, one or more functional facilities carrying out techniques herein may together form a complete software package. These functional facilities may, in alternative embodiments, be adapted to interact with other, unrelated functional facilities and / or processes, to implement a software program application, for example as a software program application such as a signal analysis facility.

[0124] Some exemplary' functional facilities have been described herein for carrying out one or more tasks. It should be appreciated, though, that the functional facilities and division of tasks described is merely illustrative of the type of functional facilities that may implement the exemplary techniques described herein, and that embodiments are not limited to being implemented in any specific number, division, or type of functional facilities. In some implementations, all functionality may be implemented in a single functional facility. It should also be appreciated that, in some implementations, some of the functional facilities described herein may be implemented together with or separately from others (i.e., as a single unit or separate units), or some of these functional facilities may not be implemented.

[0125] Computer-executable instructions implementing the techniques described herein (when implemented as one or more functional facilities or in any other manner) may, in some embodiments, be encoded on one or more computer-readable mediate provide functionality to the media. Computer-readable media include magnetic media such as a hard disk drive, optical mediaPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 such as a Compact Disk (CD) or a Digital Versatile Disk (DVD), a persistent or non-persistent solid-state memory (e g., Flash memory, Magnetic RAM, etc.), or any other suitable storage media. Such a computer-readable medium may be implemented in any suitable manner, including as computer-readable storage media 1106 of FIG. 3 described below (i.e., as a portion of a computing device 1100) or as a stand-alone, separate storage medium. As used herein, “computer-readable media” (also called “computer-readable storage media”) refers to tangible storage media. Tangible storage media are non-transitory and have at least one physical, structural component. In a “computer-readable medium,” as used herein, at least one physical, structural component has at least one physical property that may be altered in some way during a process of creating the medium with embedded information, a process of recording information thereon, or any other process of encoding the medium with information. For example, a magnetization state of a portion of a physical structure of a computer-readable medium may be altered during a recording process.

[0126] In some, but not all, implementations in which the techniques may be embodied as computer-executable instructions, these instructions may be executed on one or more suitable computing device(s) operating in any suitable computer system, including the exemplary computer system of FIG. 3, or one or more computing devices (or one or more processors of one or more computing devices) may be programmed to execute the computer-executable instructions. A computing device or processor may be programmed to execute instructions when the instructions are stored in a manner accessible to the computing device or processor, such as in a data store (e.g., an on-chip cache or instruction register, a computer-readable storage medium accessible via a bus, a computer-readable storage medium accessible via one or more networks and accessible by the device / processor, etc.). Functional facilities including these computer-executable instructions may be integrated with and direct the operation of a single multi-purpose programmable digital computing device, a coordinated system of two or more multi-purpose computing device sharing processing power and jointly carrying out the techniques described herein, a single computing device or coordinated system of computing devices (co-located or geographically distributed) dedicated to executing the techniques described herein, one or more Field-Programmable Gate Arrays (FPGAs) for carrying out the techniques described herein, or any other suitable system.

[0127] FIG. 3 illustrates one exemplary' implementation of a computing device in the form of a computing device 1100 that may be used in a system implementing techniques described herein, although others are possible. It should be appreciated that FIG. 3 is intended neither to be a depiction of necessary components for a computing device to execute a pECG analysis facility in accordance with the principles described herein, nor a comprehensive depiction.

[0128] Computing device 1100 may include at least one processor 1101 , a network adapter 1102, and computer-readable storage media 1103. Computing device 1100 may be, for example, aPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 desktop or laptop personal computer, a personal digital assistant (PDA), a smart mobile phone, a server, a wireless access point or other networking element, or any other suitable computing device. Network adapter 1102 may be any suitable hardware and / or software to enable the computing device 1100 to communicate wired and / or wirelessly with any other suitable computing device over any suitable computing network. The computing network may include wireless access points, switches, routers, gateways, and / or other networking equipment as well as any suitable wired and / or wireless communication medium or media for exchanging data between two or more computers, including the Internet. Computer-readable media 1103 may be adapted to store data to be processed and / or instructions to be executed by processor 1101. Processor 1101 enables processing of data and execution of instructions. The data and instructions may be stored on the computer-readable storage media 1103.

[0129] The data and instructions stored on computer-readable storage media 1103 may include computer-executable instructions implementing techniques which operate according to the principles described herein. In the example of FIG. 3, computer-readable storage media 1103 stores computer-executable instructions implementing various facilities and storing various information as described above. Computer-readable storage media 1103 may store pECG data analysis facility 116.

[0130] While not illustrated in FIG. 3, a computing device may additionally have one or more components and peripherals, including input and output devices. These devices can be used, among other things, to present a user interface. Examples of output devices that can be used to provide a user interface include printers or display screens for visual presentation of output and speakers or other sound generating devices for audible presentation of output. Examples of input devices that can be used for a user interface include keyboards, and pointing devices, such as mice, touch pads, and digitizing tablets. As another example, a computing device may receive input information through speech recognition or in other audible format.Examples

[0131] Described below are examples of ways in which techniques described herein may be implemented. It should be appreciated that these examples are merely illustrative, that embodiments are not limited to operating in accordance with the specific examples shown in the figures and discussed below, and that other embodiments are possible.Example 1: Predicting Left Ventricular Dysfunction and Remodeling

[0132] The purpose of this example is to provide exemplary data that may be used for training a model described herein for predicting left ventricular dysfunction and remodeling.Study Population and Patient AssignmentPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0133] Inclusion criteria for exemplary data consisted of children < 18 years old with at least one echo. Echos performed in the operating room, medical intensive care unit, or cardiac intensive care unit were excluded. Patients with known major congenital heart disease or implantable cardioverter-defibrillator / pacemaker were excluded. Patients with known major congenital heart disease based on a Fyler coding system were identified.

[0134] Each qualifying echo event was paired to an ECG; only ECG-echo pairs < 2 days apart were included. For patients with multiple ECGs within this timeframe, only the ECG closest in time to the echo was included. ECG-echo pairs with ECGs failing to pass quality control (see “Data Processing, Quality Control, and Filtering” for details) were removed. The remaining ECG- echo pairs were included as the main cohort.

[0135] A group stratified design was implemented for partitioning of the main cohort. Each patient was treated as a separate group, which restricts ECG-echo pairs for a given patient to either training or testing datasets in order to minimize leakage of ECG-echo pair data. If an ECG or echo within an ECG-echo pair was performed in the emergency department, then the ECG-echo pair was placed in the external setting group. These same patients with other ECG-echo pairs were forced into the internal testing group to ensure no data leakage occurred between training and testing. The remaining patients were then randomly partitioned 80:20 into training and internal testing datasets.

[0136] Of the 272,221 echos from 104,508 children < 18 years old without congenital heart disease, there were 122,757 ECG-echo pairs < 2 days apart. Of these ECG-echo pairs, 119,787 ECGs (61,722 patients) passed quality control, thus forming the main study cohort (FIG. 4).

[0137] The training cohort comprised of 92,377 ECG-echo pairs (46,261 patients; median age 8.2 [IQR, 2.9-13.8] years; 54% male), 8.2% with composite LV outcomes, 2.4% with LV dysfunction, 3.5% with LV hypertrophy, and 3.8% with LV dilation (TABLE 1). The internal testing cohort comprised of 24,343 ECG-echo pairs (12,631 patients; median age 8.3 [IQR, 3.0- 13.8] years; 54% male), 7.7% with composite LV outcomes, 2.4% with LV dysfunction, 3.2% with LV hypertrophy, and 3.5% with LV dilation (TABLE 1). The external testing cohort comprised of 3,067 ECG-echo pairs (2,830 patients; median age 8.0 [IQR, 1.4-14.6] years; 57% male), 10.0% with composite LV outcomes, 5.0% with LV dysfunction, 3.5% with LV hypertrophy, and 4.4% with LV dilation (TABLE 1). Further details on demographics, ECG, and echo characteristics are summarized in TABLE 1.

[0138] TABLE 1: Comparison of Demographics, ECG and Echo Characteristics, and Outcomes Stratified by Study CohortsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT

[0139] Data presented as median (interquartile range).

[0140] Abbreviations: left ventricle (LV); beats per minute (BPM); ejection fraction (EF).

[0141] Patient characteristics and outcomes stratified by age group are shown in TABLE2. ECG characteristics stratified by age are within range of previously reported values for healthy children, and confidence intervals for echo z-scores include zero. ECG findings by age include a more rightward QRS, T, and P axis for age < 1, an increasing PR and QT interval with age, and decreasing heart rate with age (TABLE 2).

[0142] TABLE 2: Comparison of Demographics, ECG and Echo Characteristics, andOutcomes Stratified by Study CohortsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCTPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0143] Data presented as median (interquartile range). Age in years.

[0144] Abbreviations: left ventricle (LV); beats per minute (BPM); ejection fraction (EF).

[0145] TABLES 3-6 highlight the numerous significant differences in ECG-echo pair demographics, ECG characteristics, and echo data when stratifying by each outcome. Of note, approximately 40% of patients with LV dysfunction had concomitant LV dilation (TABLE 4), approximately 10% of patients with LV hypertrophy had concomitant LV dilation (TABLE 5). Approximately 30% and 10% of patients with LV dilation also had LV dysfunction and hypertrophy, respectively (TABLE 6).

[0146] TABLE 3: Comparison of ECG-Echo Pair Demographics, ECG Characteristics, and Echo Characteristics Stratified by Composite OutcomePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0147] Data presented as median (interquartile range). Age in years. P-values obtained by Wilcoxon rank sum test, Fisher's exact test, or Pearson's Chi-squared test, as appropriate.

[0148] Abbreviations: left ventricle (LV); beats per minute (BPM); ejection fraction (EF).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0149] TABLE 4: Comparison of ECG-Echo Pair Demographics, ECG Characteristics, and Echo Characteristics Stratified by LV Dysfunction OutcomePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0150] Data presented as median (interquartile range). Age in years. P-values obtained by Wilcoxon rank sum test, Fisher's exact test, or Pearson's Chi-squared test, as appropriate.

[0151] Abbreviations: left ventricle (LV); beats per minute (BPM); ejection fraction (EF).

[0152] TABLE 5: Comparison of ECG-Echo Pair Demographics, ECG Characteristics, and Echo Characteristics Stratified by LV Hypertrophy OutcomePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0153] Data presented as median (interquartile range). Age in years. P-values obtained by Wilcoxon rank sum test, Fisher's exact test, or Pearson's Chi-squared test, as appropriate.

[0154] Abbreviations: left ventricle (LV); beats per minute (BPM); ejection fraction (EF).

[0155] TABLE 6: Comparison of ECG-Echo Pair Demographics, ECG Characteristics, and Echo Characteristics Stratified by LV Dilation OutcomePCT / US25 / 48621 30 September 2025 (30.09.2025)CT 025PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0156] Data presented as median (interquartile range). Age in years. P-values obtained by Wilcoxon rank sum test, Fisher's exact test, or Pearson's Chi-squared test, as appropriate.

[0157] Abbreviations: left ventricle (LV); beats per minute (BPM); ejection fraction (EF). Data Retrieval

[0158] Raw ECG signals were exported from a MUSE ECG data management system (GE Healthcare, Chicago, IL). Waveform data were obtained from XML files, where each onedimensional vector of data sampled at a rate of 250 Hz for 10 seconds duration (2500 samples) corresponds to a lead (I, II, and V1-V6). Linear transformations of the vectors were performed based on the Einthoven law and Goldberger equation to obtain leads III, aVF, aVL, and aVR. Age, sex, and phy sician reviewed ECGs measurements (e.g., QRS interval, QRS axis, T axis, P axis, PR interval, QT interval, QTc, and heart rate) are archived in an internal database at Boston Children’s Hospital, which was also retrieved. In addition, ECG-based diagnoses of LV hy pertrophy (as coded by expert pediatric electrophysiologists) were retrieved for benchmarking purposes.

[0159] Similarly, echo reports written by pediatric cardiologists are archived in an internal database at Boston Children’s Hospital; extracted records contained the human expert classification of the degree of left ventricular (LV) systolic dysfunction, hypertrophy, and / or dilation (if any). Potential grades were “trivial”, “mild”, “mild-to-moderate”, “moderate”, “moderate-to-severe”, and “severe”. When available, quantitative measures of LV ejection fraction (% and z-score), LV mass (raw and z-score), LV mass / volume (raw and z- score), and LV end- diastolic volume (raw and z-score) were obtained. For both extracted ECG and echo records, unique patient identifiers and dates and times are available to link ECG-echo pairs and filter based on timing of events.Quality Control and Data Preprocessing

[0160] The purpose of this example is to provide an exemplary method of preprocessing the ECG data. In the case of multiple ECG recording attempts for a given ECG event, the final recorded ECG is retrieved. This ECG is then discarded if any lead is not 2500 samples long, or ifPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 any lead recording has no lead information (i.e., flat line). Given that ECGs are prone to recording errors (e.g., baseline wander; electrical interference), a high pass filter was utilized. The ECG was then trimmed to 2048 samples (approximately 8 seconds) to facilitate conveniently working with convolution neural networks.Definition of Primary Outcomes

[0161] Individual outcomes, i.e., exemplary cardiac conditions used fortraining the models described herein, included LV systolic dysfunction, LV hypertrophy, and LV dilation. Human expert knowledge was considered as a ground truth, whereby: 1) LV systolic dysfunction was considered positive if the echo report was coded by a pediatric cardiologist for qualitatively greater than “mild” LV systolic dysfunction; 2) LV hypertrophy was considered positive if the echo report was coded by a pediatric cardiologist for qualitatively greater than “mild” LV hypertrophy, or LV hypertrophic cardiomyopathy; and 3) LV dilation was considered positive if the echo report was coded by a pediatric cardiologist for qualitatively greater than “mild” LV dilation, or LV dilated cardiomyopathy. The composite outcome was defined as having positive LV systolic dysfunction, hypertrophy, or dilation. The primary outcomes were used to train the models used herein.

[0162] As a secondary subgroup performance analysis of the human expert trained model, quantitative cutoffs were implemented for the above outcomes, whereby LV ejection fraction, LV mass, and LV end-diastolic volume z-scores of < -2.5, > 2.5, and > 2.5 (corresponding to quantitative “mild” cutoffs) were considered positive for LV dysfunction, hypertrophy, and dilation, respectively. LV ejection fraction, LV mass, and LV end-diastolic volume z-score cutoffs of < -4.0, > 4.0, and > 4.0 (corresponding to quantitative “moderate” cutoffs), respectively, as well as < -6.0, > 6.0, and > 6.0 (corresponding to quantitative “severe” cutoffs), respectively, were also considered.Model Selection, Architecture, and Training

[0163] An exemplary model was developed solely on the training set, which was further partitioned 95% for training and 5% for validation to allow for hyperparameter tuning. A convolutional neural network used 12 x 2048 ECG inputs in a convolutional neural network similar to a residual network (i.e., including skip connections) that is adapted for uni dimensional signals.

[0164] An exemplary artificial intelligence-enhanced pediatric ECG (AI-pECG) network consisted of a convolutional layer followed by four residual blocks with two convolutional layers per block.

[0165] The output of each convolutional layer is rescaled using batch normalization, and fed into a rectified linear activation unit, after which dropout at a rate of 0.8 is applied. Max pooling and convolutional layers with filter length 1 are included in the skip connections to make the dimensions match those from the signals in the main branch. The output of the last block is fedPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 into a fully connected layer with a sigmoid activation function. Demographics (i.e., age and sex) were also incorporated as inputs along with ECG waveforms in a separate deep learning model (AI-pECG + age + sex). More specifically, the above architecture was modified by adding a separate part of the model, where demographics (i.e., age and sex) were concatenated and pass it through a fully connected layer. The outputs for demographic and ECG model parts are individually flattened and then concatenated to obtain one feature vector. The resulting feature vector was fed into the final fully connected layer with a sigmoid activation function.

[0166] For each model, the final configuration hyperparameters were obtained via a grid search on the training set among the following options: kernel size [3, 9, 17], batch size [8, 32, 64], and initial learning rate [0.01, 0.001, 0.0001, 0.00001], The average cross-entropy was minimized using the Adam optimizer. Maximum 150 epochs with early stopping were used based on validation loss. The model with the lowest validation loss during hyperparameter tuning was selected as an exemplary final trained model.Performance Evaluation and Statistical Analyses

[0167] Model performance was evaluated on the internal and external test groups. Given the nature of an imbalanced dataset, the area under the receiver operating curve (AUROC) as well as the area under the precision-recall (i.e., positive predictive value-sensitivity) curve (AUPRC) were computed. The DeLong test was performed to compare AUROCs across models. To benchmark an LV hypertrophy model, pediatric electrophysiologist expert ECG-based diagnoses of LV hypertrophy (using the first available ECG per patient to minimize human bias from prior echo findings) were used. Other performance metrics evaluated included positive predictive value, negative predictive value, sensitivity, and specificity. For all metrics, a higher value is indicative of better performance. Resampling with 1,000 bootstraps was implemented to obtain performance metric confidence intervals.

[0168] After training an exemplary' AI-pECG model on nearly 100,000 ECG-echo pairs with corresponding human expert classified greater than mild LV dysfunction, hypertrophy, and dilation, model performance was tested.

[0169] During internal testing, the exemplary AI-pECG model achieved AUROCs of 0.85 [95% CI, 0.84-0.86], 0.88 [95% CI, 0.86-0.89], 0.85 [95% CI, 0.83-0.86] and 0.86 [95% CI, 0.84- 0.87] for the LV composite outcome, LV dysfunction, LV hypertrophy, and LV dilation, respectively (FIG. 5). Notably, the AI-pECG model outperformed the pediatric electrophysiologist expert ECG-based diagnosis of LV hypertrophy, which had a sensitivity' of 33%, specificity of 95%, and positive predictive value of 11.1% (FIG. 5; grey dot). Adding age and sex to the AI- pECG model led to similar performance for the LV composite outcome (p=0.07), LV hypertrophy (p=0.3), and LV dilation (p=0.3), with a minor yet statistically significant difference for LVPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 dysfunction (AUROC 0.86 [95% CI 0.85-0.88]; p<0.01) (FIG. 5). When looking at single random ECGs per patient, model performance remained high (FIG. 6).

[0170] During testing on the external cohort, the AI-ECG model achieved AUROCs of 0.81 [95% CI, 0.78-0.84], 0.83 [95% CI, 0.79-0.87], 0.84 [95% CI, 0.80-0.87] and 0.86 [95% CI, 0.83-0.90] for the LV composite outcome, LV dysfunction, LV hypertrophy, and LV dilation, respectively (FIG. 7). The AI-pECG model again outperformed the pediatric electrophysiologist expert ECG-based diagnosis of LV hypertrophy, which had a sensitivity of 24%, specificity of 95%, and positive predictive value of 13.2% (FIG. 7; grey dot). Adding age and sex to the AI- ECG model led to similar performance for the LV composite outcome (p=0.8), LV dysfunction (p=0.4), LV hypertrophy (p=0.07), and LV dilation (p=0.9). (FIG. 7). When looking at single random ECGs per patient, model performance remained high (FIG. 8).

[0171] Performance of the AI-pECG model, trained on human expert qualitative cutoffs, was next explored to discriminate between quantitative cutoffs for LV dysfunction (LV ejection fraction z-score < -2.5, -4, -6.0), LV hypertrophy (LV mass z-score > +2.5, +4, +6.0), LV dilation (LV end-diastolic volume z-score > +2.5, +4, +6.0), and the composite outcome for each corresponding cutoff (FIG. 9). In general, performance when using a z-score cutoff of 2.5 was comparable to performance using qualitative cutoffs (FIG. 9). Performance increased with a higher cutoff deviating from normal, with AUROCs of > 0.85 and > 0.9 for every outcome in both internal and external test groups using a z-score cutoff of 4.0 and 6.0, respectively.Subgroup Analyses

[0172] Age and sex are known to influence ECG characteristics in a healthy pediatric population, and were therefore explored in subgroup analyses. Age partitioning was adapted with groupings of age < 1, 1 < age < 3, 3 < age < 8, 8 < age < 12, and 12 < age < 18 years. AUROCs were calculated for each subgroup.AI-pECG performance when stratifying by age and sex was next examined (FIG. 10). In the LV composite, LV dysfunction, and LV hypertrophy groups, performance varied with age, most notably for predicting LV hypertrophy in age < 1 year. The algorithm appeared to have slightly better performance in females compared to males for predicting the LV composite outcome, LV dysfunction, and LV dilation (FIG. 10).Saliency Mapping

[0173] Saliency mapping may be used for model interpretation using the methods described herein. In an effort to provide model interpretability, the following analyses were performed to identify which features of the ECG input may contribute to model prediction: 1) median waveform analysis; and 2) saliency mapping. Similar to others, median waveform analysis is a technique for visualizing aggregated ECG samples into a single beat. By doing so, examplesPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 of high-risk and low-risk ECGs may be visualized. Herein, the 100 highest predicted ECGs for a given outcome in the internal test set were used to create high-risk median waveforms, and the 100 lowest predicted ECGs in the internal test set were used to create low-risk median waveforms. Median waveforms in each lead were generated by: 1) QRS complex detection; 2) interpolating all ECGs to the same heart rate; 3) computing the median voltage across beats for each patient; 4) computing the median voltage across patients for each time bin in the cardiac cycle. Saliency mapping was helpful in identifying which features of the ECG input contribute to model prediction. Saliency maps highlight components of the ECG where a change in input (i.e., ECG voltage) leads to a change in prediction. Saliency maps were created using a Shapley Additive Explanations (SHAP) framework. To highlight the most influential components of the ECG waveform, the SHAP values for the high-risk ECGs were obtained. Subsequently, the above steps to generate median waveforms were implemented on SHAP values over time. The resultant darker regions in saliency maps correspond to greater contribution to the prediction.

[0174] As shown in FIG. 11, to predict LV hypertrophy, the salient features in some cases may be or include precordial QRS complexes. High-risk features to predict LV hypertrophy include deep S waves in V1-V2. In limb lead I, the QRS complex was also salient, with high-risk features including a high amplitude R wave. For LV dilation, the salient features in some cases may be or include lateral precordial (V4-V6) QRS complexes. High-risk features to predict LV dilation include high amplitude R waves in V4-V6. As shown in FIG. 12, the saliency map for the composite outcome appears to merge features from each of the individual outcomes of interest.Example 2: Deep Learning-Based Electrocardiogram Analysis Predicts Biventricular Dysfunction and Dilation in Congenital Heart Disease

[0175] The purpose of this example is to provide exemplary data and uses cases of a trained model described herein.

[0176] Patients with congenital heart disease (CHD) are a heterogeneous population with complex anatomy and physiology. In the current era, the majority survive to adulthood.Noninvasive imaging plays a central role in risk stratifying this growing population and informing medical / surgical interventions, with biventricular size and function among the strongest predictors of long-term mortality in a multitude of CHD lesions. For example, in adults with tetralogy of Fallot (ToF), left ventricular (LV) and right ventricular (RV) dysfunction are both predictive of death and sustained ventricular tachycardia.

[0177] Cardiac magnetic resonance (CMR) imaging has an established role in the lifelong management of patients with CHD to accurately assess LV, RV, and functional single ventriclePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 size and ejection fraction (EF), which are particularly more challenging to measure by echocardiography. However, CMR has practical limitations prohibiting its widespread use including being time-, resource-, and cost-intensive, making it of interest to develop a cheap and effective tool to help inform timing of CMR.

[0178] Electrocardiograms (ECGs) are a quick, ubiquitous, and cost-effective tool used for cardiac screening of adults and children. Artificial intelligence-enhanced electrocardiogram (AI-ECG) algorithms reliably predict a range of cardiovascular phenotypes in the general adult population, including biventricular dilation and dysfunction. However, there remains a paucity of AI-ECG applications for the CHD population with distinct ECG characteristics attributed to age, CHD lesion, and prior interventions. AI-ECG technology has yet to be applied to predict the gold-standard CMR measurements that congenital cardiologists rely on to risk stratify and inform management of CHD patients.

[0179] In this study, a primary objective was to address this gap by developing, internally testing, and externally validating an AI-ECG model to predict LV and RV dysfunction and dilation in patients with and without CHD (FIG. 13A-13C).

[0180] FIGS. 13A-C Artificial Intelligence-Enhanced Electrocardiography to Predict Biventricular Size and Function

[0181] An artificial intelligence (Al)-enhanced electrocardiography (ECG) algorithm trained on ECG- cardiovascular magnetic resonance (CMR) pairs at Boston Children’s Hospital was predictive of right ventricular (RV) and left ventricular (LV) dysfunction and dilation in a congenital heart disease cohort, with external validation and model explainability. EDV = end- diastolic volume; EF = ejection fraction.METHODS

[0182] Internal study population and patient assignment

[0183] We utilized patient data from Boston Children’s Hospital between 2002 and 2021. All CMR studies with RV and LV EF percentage and end-diastolic volume (EDV) z-scores were considered eligible. LVEDV and RVEDV z-scores were calculated using published equations; patients with body surface area <1.0 m2 were excluded given the paucity of normative CMR data for this group, as well as the phenomenon of heteroscedasticity preventing extrapolation. Each qualifying CMR event was paired with an ECG; only ECG-CMR pairs <30 days apart without an intermediate catheterization or surgery were included. In cases of multiple ECGs within this timeframe, only the ECG closest in time to the CMR was included. ECG-CMR pairs with ECGs failing to pass quality control (data not provided) were removed. The remaining ECG-CMR pairs were included as the main internal cohort (FIG. 14).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0184] FIG. 14 STROBE Diagram Showing Initial Patient Selection and the FinalCohort

[0185] Filtering at each data processing stage is shown. Patient partitioning for training (80%) and testing (20%) is shown, with external validation at Mount Sinai. CMR = cardiovascular magnetic resonance; EDV = end-diastolic volume; EF = ejection fraction; LV = left ventricle; QC = quality control; RV = right ventricle.

[0186] To assign patients into CHD subgroups, an institutional Fyler coding system was utilized. The coding system allows for identification of patients with specific structural diagnoses (e.g., coarctation of the aorta, ventricular septal defect, and so on) as well as clinical diagnoses (e.g., myocarditis, right heart failure, and so on). Based on the primary underling cardiac diagnosis, patients were grouped into 4 categories: 1) functionally single ventricles at any stage of palliation with reportable LV and RV function and size (inclusive of hypoplastic left heart syndrome, tricuspid atresia, double outlet LV, double inlet LV, double outlet RV, double inlet RV; i.e., functionally single-ventricle patients without a secondary ventricle were excluded); 2) RV at-risk including ToF, right heart failure, right-dominant atrioventricular canal defect, atrial septal defect, pulmonary atresia, total anomalous pulmonary venous return, Ebstein anomaly, and truncus arteriosus; 3) LV at-risk including coarctation of the aorta, left heart failure, myocardial infarction, left-dominant atrioventricular canal defect, L- or D-loop transposition of the great arteries, anomalous left coronary artery from the pulmonary artery, cardiomyopathy, heart transplant, ventricular septal defect, myocarditis, anemia, and iron overload; and 4) other. Grouping was tiered, such that if a patient met functionally single ventricle criteria, the patient was excluded from the LV or RV at-risk group. Next, if a patient met RV at-risk criteria, the patient was excluded from the LV at-risk group. Finally, if a patient did not meet any group criteria, the patient was placed in the other group. Note the other group includes non-CHD to reflect all indications of CMR at the population institution and diversify the training set. Patient diagnoses within each subgroup are shown in TABLE 7.

[0187] TABLE 7: Patient Diagnoses with Each CMR Subgroup„ . . LV Group RV Group SV Group OtherDiagnosis N = 2,639 pairs N = 2,996 pairs N = 822 pairs N = 2,127 pairsSingle ventricle 0 (0%) 0 (0%) 378 (46%) 0 (0%)HLHS 0 (0%) 0 (0%) 211 (26%) 0 (0%)Fontan 0 (0%) 0 (0%) 375 (46%) 0 (0%)Glenn 0 (0%) 0 (0%) 175 (21%) 0 (0%)Tricuspid Atresia 0 (0%) 0 (0%) 151 (18%) 0 (0%)DOLV 0 (0%) 0 (0%) 13 (1.6%) 0 (0%)DILV 0 (0%) 0 (0%) 193 (23%) 0 (0%)DORV 0 (0%) 0 (0%) 381 (46%) 0 (0%)DIRV 0 (0%) 0 (0%) 9 (1.1%) 0 (0%)PCT / US25 / 4862130 September 2025 (30.09.2025)Atorney Docket No.: 167705-037401 / PCTElectronic Deposit Date: September 30, 2025Left Heart Failure 32(1.2%) 26(0.9%) 6(0.7%) 0(0%)Coarctation of the785(30o / o)97(3.2%) 212(26%) 0(0%) / \.onaMyocardial 74(2.8%) 55(1.8%) 21 (2.6%) 0(0%)InfarctionL-dominant CAVC 0(0%) 2(<0.1%) 14(1.7%) 0(0%)Balanced CAVC 17(0.6%) 10(0.3%) 15(1.8%) 0(0%)L-loopTGA 137(5.2%) 72(2.4%) 211 (26%) 0(0%)D-loopTGA 554 (21%) 191 (6.4%) 283 (34%) 0(0%)ALCAPA 26(1.0%) 16(0.5%) 0(0%) 0(0%)^ny..+, 1,027 (39%) 271 (9.0%) 119(14%) 0(0%)CardiomyopathyHeart Transplant 3(0.1%) 0(0%) 0(0%) 0(0%)VSD 686 (26%) 935 (31%) 469(57%) 0(0%)Myocarditis 114(4.3%) 8(0.3%) 5(0.6%) 0(0%)Iron Overload 104(3.9%) l(<0.1%) 0(0%) 0(0%)Tetralogy of Fallot 0(0%) 1,985 (66%) 167 (20%) 0(0%)Right Heart Failure 0(0%) 188(6.3%) 59(7.2%) 0(0%)R-dominant CAVC 0(0%) 2(0.1%) 40(4.9%) 0(0%)Pulmonary Atresia 0(0%) 664(22%) 218(27%) 0(0%)TAPVR 0 (0%) 79 (2.6%) 28 (3.4%) 0 (0%)Ebstein anomaly 0(0%) 336 (11%) 35(4.3%) 0(0%)Truncus Arteriosus 0(0%) 163(5.4%) 16(1.9%) 0(0%)

[0188] A group stratified design was implemented for partitioning of the main cohort.ECG-CMR pairs for a given patient were restricted to either training or testing data sets to minimize leakage of ECG-CMR pair data. The patients were randomly partitioned 80:20 into training and testing data sets.

[0189] External study population

[0190] The external validation cohort (Mount Sinai Hospital, New Y ork, New Y ork, USA) had similar inclusion criteria such that only ECG-CMR pairs <30 days apart without an intermediate procedure were included.

[0191] Data Retrieval

[0192] All internal raw ECG signals were exported from the MUSE ECG data management system (GE Healthcare, Chicago, IL). Waveform data were obtained from XML files, where each one-dimensional vector of data sampled at a rate of 250 Hz for 10 seconds duration (2500 samples) corresponds to a lead (I, II, and V1-V6). External data was resampled to similarly have a sampling rate of 250 Hz for 10 seconds duration. Linear transformations of the vectors were performed based on the Einthoven law and Goldberger equation to obtain leads III, aVF, aVL, and aVR. Age, sex, and physician reviewed ECGs measurements (e.g., QRS interval, QRS axis, T axis, P axis, PR interval, QT interval, QTc, and heart rate) are archived in an internal database at Boston Children’s Hospital, which were also retrieved. CMR measurementsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 including LVEF, LV end-diastolic volume (LVEDV) (raw and z-score), RVEF, and RVEDV (raw and z-score) were extracted from the departmental database.

[0193] Quality Control and Data Preprocessing

[0194] In the case of multiple ECG recording attempts for a given ECG event, the final recorded ECG is retrieved. This ECG is then discarded if any lead is not 2500 samples long, or if any lead recording has no lead information (i.e. , flat line). Given that ECGs are prone to recording errors (e.g., baseline wander; electrical interference), a high pass filter was utilized with cutoff frequency 0.8 Hz, rejection band 0.2 Hz, ripple in passband 0.5 dB, and attenuation in rejection band 40 dB. The ECG was then trimmed to 2048 samples (approximately 8 seconds) to facilitate conveniently working with convolution neural networks.

[0195] Subgroup Analysis

[0196] Age and sex are known to influence ECG characteristics in a healthy pediatric population. Age partitioning was grouped as age < 20, 20 < age < 40, age > 40 years. In addition, ECG characteristics are variable across different CHD lesions. Therefore, populations were partitioned based on risk groupings. AUROCs and AUPRCs were calculated for each subgroup.

[0197] Definition of outcomes

[0198] The individual outcomes included greater than mild LV dysfunction (LVEF <40%), RV dysfunction (RVEF <35%), LV dilation (LVEDV z-score >4, corresponding to 121 mL / m2 in women and 141 mL / m2 in men), and RV dilation (RVEDV z-score >4, corresponding to 130 mL / m2 in women and 143 mL / m2 in men). The composite biventricular dysfunction outcome was defined as LVEF <40% and RVEF <35%. The primary outcomes were used to train and test the model used herein. In a secondary RV volume-specific model, individual outcomes included RVEDV z-score >4, indexed RVEDV >160 mL / m2, and indexed RVEDV >180 mL / m2.

[0199] As a secondary outcome analysis, time to all-cause mortality after CMR was evaluated. When multiple ECG-CMR pairs were available in a patient, a randomly selected ECG-CMR pair was included for survival analysis.

[0200] Model selection, architecture, and training

[0201] A transfer learning approach was utilized for model development. Model weights from an AI-ECG model were used to predict left ventricular dysfunction or remodeling in patients <18 years of age without major CHD, and then trained the model on the aforementioned training set.

[0202] The training set was further partitioned 95% for training and 5% for validation to allow for hyperparameter tuning. A convolutional neural network used 12 x 2,048 ECG inputsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 with an architecture inspired by the residual network (i.e., including skip connections) adapted for uni dimensional signals. Network architecture was similar to the inventor’s previous work.

[0203] Model hyperparameters were tuned by performing a grid search on the training set over the following values: kernel size [3, 9, 17], batch size [8, 32, 64], and initial learning rate [0.01, 0.001, 0.0001], The average cross-entropy was minimized using the Adam optimizer. A maximum 150 epochs were used with early stopping based on validation loss. The model with the lowest validation loss during hyperparameter tuning was selected as the final model (kernel size 17, batch size 32, learning rate 0.001).

[0204] Performance evaluation

[0205] Model performance was evaluated only on the test group. To account for class imbalance and capture model performance at various thresholds, the area under the receiver operating curve (AUROC) and area under the precision-recall (i.e., positive predictive valuesensitivity) curve (AUPRC) were evaluated. In addition, positive predictive value, negative predictive value, sensitivity, and specificity were evaluated at the Youden index threshold (i.e., maximizing sensitivity and specificity) in the training set. Metric Cis were computing using 1,000 bootstrap resamples.

[0206] Survival analysis

[0207] For mortality analysis, Kaplan-Meier curves were constructed to visualize survival probabilities of patients based on their AI-ECG predictions. Patients were deemed high- or low- risk using the Youden Index in the training cohort as a cutoff threshold. Kaplan-Meier curves were generated using a single random ECG-CMR pair per patient. Patients were censored at the time last known alive. Statistical comparison between Kaplan-Meier curves was performed using the log-rank test.

[0208] Model explainability

[0209] To explain model behavior across all outcomes, the following analyses were performed: 1) median waveform analysis; and 2) saliency mapping. Median waveform analysis is a technique to visualize aggregated ECG samples into a single beat. In doing so, examples of high-risk and low-risk ECGs can be visualized. Herein, the 25 highest predicted ECGs for a given outcome were used in the internal test set to create high-risk median waveforms, and the 25 lowest predicted ECGs in the internal test set to create low-risk median waveforms. Median waveforms were generated in each lead using the NeuroKit Python toolbox by: 1) QRS complex detection; 2) interpolating all ECGs to the same heart rate; 3) computing the median voltage across beats for each patient; 4) computing the median voltage across patients for each time bin in the cardiac cycle.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0210] Saliency mapping helps identify which features of the ECG input contribute to model prediction. Saliency maps highlight components of the ECG where a change in input (i.e., ECG voltage) leads to a change in prediction. Saliency maps were created using a Shapley Additive Explanations (SHAP) framework. To highlight the most influential components of the ECG waveform, the SHAP values for the high-risk ECGs were obtained. Subsequently, the above steps to generate median waveforms were implemented on SHAP values over time. The resultant darker regions in saliency maps correspond to greater contribution to the prediction.

[0211] Data availability and software

[0212] Requests for Boston Children’s Hospital data and related materials will be internally reviewed to clarify if the request is subject to intellectual property or confidentiality constraints. Shareable data and materials will be released under a material transfer agreement for noncommercial research purposes. Institutional Review Board approval was obtained by each respective institution in this study.

[0213] Programming Code

[0214] The convolutional neural network used the Keras framework with a Tensorflow (Google Inc, Mountain view, CA) backend using Python 3.9. Deep learning was executed on institutional graphics processing units. All other pre- and post-processing code was written in Python 3.9 and R 4.0, which was executed locally.RESULTS

[0215] Patient population baseline characteristics and outcomes

[0216] Of the 18,526 ECG-CMR pairs <30 days apart, 12,859 were without an intermediate surgery or catheterization, with 8,701 ECGs (n = 4,993) meeting entry criteria and 8,584 ECGs (n = 4,941) passing quality' control, thus forming the main study cohort (FIG. 14).

[0217] The training cohort comprised of 6,833 ECG-CMR pairs (n = 3,954; median age at CMR 20.7 years [Q1-Q3: 15.5-30.4 years]; 56% men; 19% with ToF), 3.4% with LV dysfunction, 4.4% with RV dysfunction, 8.3% with LV dilation, and 18% with RV dilation (TABLE 8). A majority' of the patients were in the LV (31%) and RV at-risk groups (34%), followed by the other (25%) and single-ventricle (10%) groups. When stratifying by group, there were notable differences in ECG, CMR, and outcome characteristics (TABLE 9). Further details on demographics, ECG, and CMR characteristics are summarized in TABLE 8, with similar baseline characteristics in the internal test group. Within the training and test cohorts, 4.4% of patients died at a median age of 29.6 years (Q1-Q3: 20.7-45.2 years) and 24. 1 years (Q1-Q3: 18.3-43.2 years), respectively.

[0218] TABLE 8: Baseline Characteristics of Internal Training and Testing CohortsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0219] Values are n, n (%), or median (Q1-Q3).

[0220] CMR = cardiovascular magnetic resonance; ECG = electrocardiogram; EDV = end-diastolic volume; EF = ejection fraction; LV = left ventricular; RV = right ventricular; SV = single ventricle.

[0221] TABLE 9: Baseline ECG-CMR Pair Characteristics Stratified by CMR GroupPCT / US25 / 4862130 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025LV Group RV Group SV Group OtherN = 2,639 pairs N = 2,996 pairs N = 822 pairs N = 2, 127 pairsAge at CMR (years) 20.1 (15.4,290) 22.1 (15.7,32.8) 21.5 (160,30.1) 19.7(154,27.7)ECG-CMR PairCharacteristicHeart Rate (bpm) 70.0(60.0,810) 72.0(63.0,82.0) 72.0(630,82.0) 69.0(610,79.0)QRS axis 70.0 (48.0,900) 84.5 (64.0, 104.0) 94.0(67.0, 134.8) 69.0(510,83.0)Taxis 52.0 (34.0,710) 62.0(46.0,75.0) 76.0 (530,99.0) 52.0(370,66.0)Paxis 47.0(31.0,610) 45.0(30.0,58.0) 49.0(290,68.0) 49.0(330,62.0)PR interval (ms) 150.0 (134.0, 168.0) 156.0 (138.0, 178.0) 158.0 (136.0, 180.0) 148.0 (134.0, 166.0)QRS interval (ms) 98.0 (88.0, 112.0) 134.0 (100.0, 154.0) 112.0 (980, 136.0) 94.0 (84.0, 102.0)QT interval (ms) 396.0(371.5,424.0) 416.0(388.0,442.0) 408.0(380.0,438.0) 392.0 (370.0,415.0)QTc interval (ms) 428.0(409.0,447.0) 452.0(430.0,478.0) 445.0(424.0,470.0) 421.0(406.0,438.0)LVEF (%) 59.8 (54.7,648) 57.7(53.7,62.0) 55.8 (490,60.8) 60.4(565,64.5)RVEF (%) 56.4 (50.7,617) 51.0(45.4,56.0) 50.5 (433,57.3) 56.4(522,60.6)LVEDV(mL) 159.0(128.3,200.4) 143.7(117.2, 177.5) 134.6(942, 169.2) 162.8(127.0,209.1)LVEDVz-score 0.8 (-0.2, 2.1) 0.4 (-0.5, 1.3) 0.0 (-1.6, 1.5) 0.9 (-0.1, 2.4)RVEDV(mL) 166.3(130.4,205.8) 213.3 (166.4,266.6) 150.4(892,208.5) 169.0 (131.7,208.9)RVEDVz-score 0.9 (-0.1, 2.1) 3.2 (1.6, 5.1) 0.9 (-1.7, 2.6) 10 (0.1, 2.0)OutcomeLVEF <40% 115(4.4%) 63(21%) 100(12%) 7(0.3%)RVEF <35% 81 (3.1%) 193(6.4%) 93(11%) 18(0.8%)LVEDVz-score >4 238(9.0%) 142(4.7%) 55(6.7%) 243(11%)RVEDV z-score > 4 176(6.7%) 1,133 (38%) 126(15%) 152(7.1%)

[0222]

[0223]

[0224] The external cohort comprised of 909 ECG-CMR pairs (746 patients; median age at CMR 25.4 years [Q1-Q3: 15.8-32.3 years]; 56% men; 15% with ToF), 4.1% with LV dysfunction, 7.2% with RV dysfunction, 5.5% with LV dilation, and 14% with RV dilation.

[0225] AI-ECG model performance

[0226] Model performance (FIGS.15A-D, TABLE 7) was similar between internal testing (LV dysfunction: AUROC: 0.87, AUPRC: 0.24; LV dilation: AUROC: 0.86, AUPRC: 0.36; RV dysfunction: AUROC: 0.88, AUPRC: 0.35; RV dilation: AUROC: 0.81, AUPRC: 0.50) and external cohort (LV dysfunction: AUROC: 0.89, AUPRC: 0.35; LV dilation: AUROC: 0.83, AUPRC: 0.29; RV dysfunction: AUROC: 0.82, AUPRC: 0.31; RV dilation: AUROC: 0.80, AUPRC: 0.44). When looking at single random ECG-CMR pairs per patient, model performance remained similar (FIG.16). Model performance was comparable in detecting biventricular dysfunction (FIG.17).

[0227] FIGS.15A-D External Validation of Electrocardiogram-Based Deep Learning Model PerformancePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0228] Model performance evaluated with receiver operating and precision-recall curves on internal (left) and external (right) cohorts for: (A) left ventricular ejection fraction (LVEF) <40%; (B) left ventricular end-diastolic volume (LVEDV) z-score >4; (C) right ventricular ejection fraction (RVEF) <35%; and (D) right ventricular end-diastolic volume (RVEDV) z- score >4. Area under receiver-operating curve (AUROC) and area under precision recall curve (AUPRC) metric values for each model and outcome are inset. Dotted line represents chance. 95% Cis are shown using bootstrapping. PPV = positive predictive value.

[0229] TABLE 10: Model Performance Metrics Across Internal and External CohortsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0230] Values are median (95% CI).

[0231] AUPRC = area under the precision-recall curve; AUROC = area under the receiver operating curve; other abbreviations as in TABLE 9.

[0232] Model performance metrics were subsequently evaluated (TABLE 10) at the Youden index threshold in the training set. LVEF, LV dilation, and RV dilation had similar sensitivities (ranging from 0.78-0.94) across internal test and external cohorts. In contrast, for RVEF, the sensitivity was markedly higher in the internal test cohort compared with the external validation cohort (0.83 vs 0.53, respectively).

[0233] FIG. 16 illustrates Electrocardiogram-Based Deep Learning Model Performance on Random ECG-CMR Pairs. Model performance evaluated with receiver operating (left) and precision-recall (right) curves using single random ECG-CMRs per patient for the: (A) left ventricular ejection fraction (LVEF); (B) Left ventricular end-diastolic volume (LVEDV); (C) right ventricular ejection fraction (RVEF); and (D) Right ventricular end-diastolic volume (RVEDV). AUROC and AUPRC metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping.

[0234] Abbreviations: area under the receiver operating curve (AUROC), area under precision recall curve (AUPRC), positive predictive value (PPV)

[0235] FIG. 17 illustrates Electrocardiogram-Based Deep Learning Model Performance for Detecting Biventricular Dysfunction. Model performance evaluated with receiver operating (left) and precision-recall (right) curves for biventricular dysfunction (i.e., LVEF < 40% and RVEF < 35%). AUROC and AUPRC metnc values for each model and outcome are inset.Dotted line represents chance. 95% confidence intervals are shown using bootstrapping.

[0236] Abbreviations: positive predictive value (PPV).

[0237] Subgroup analysis

[0238] We next examined ALECG performance when stratifying by age and sex (FIG.18) (for full model performance metric details, see TABLES 11-12). There was no clear relationship between age and model performance. Performance was higher for men than women for RVEF and RVEDV.

[0239] TABLE 11: Model Performance Metrics Across Age SubgroupsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCTPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCTElectronic Deposit Date: September 30, 2025

[0241]

[0242] FIG. 18 Subgroup Model Performance

[0243] Forest plot showing AUROC (blue) and AUPRC (red) performance when stratifying by age and CMR subgroups for the following outcomes: LVEF <40%, LVEDV z- score >4, RVEF <35%, and RVEDV z-score >4. Dotted lines represent overall cohort performance, with metrics inset above each forest plot. 95% Cis are shown using bootstrapping. SV = single ventricle; other abbreviations as in FIGS. 14-15.

[0244] We similarly performed subgroup analysis by CMR grouping (FIG. 18, TABLE 13). For LVEF, the LV at-risk group performed better than the RV at-risk group, whereas for LVEDV, the RV at-risk group performed better than the LV at-risk group. Across all outcomes, the single-ventricle group had the lowest performance. For RVEDV, performance was worse for the RV at-risk group than the LV at-risk group. Overall model performance was broadly insensitive to exclusion of common CHD lesions in this cohort (FIG. 19).

[0245] FIG. 19 illustrates Sensitivity Analysis of Overall Model Performance Excluding Each Lesion. Forest plot showing area under the receiver operating (AUROC; red) and precision recall (AUPRC; black) curve performance when excluding specific lesions for the following outcomes: left ventricular ejection fraction (LVEF) < 40%, left ventricular end-diastolic volume (LVEDV) z-score > 4, right ventricular ejection fraction (RVEF) < 35%, and right ventricular end-diastolic volume (RVEDV) z-score > 4. 95% confidence intervals are shown using bootstrapping.

[0246] TABLE 13: Model Performance Metrics Across CMR Indication SubgroupsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0247] Finally, a specific CHD lesion was assessed with a common CMR indication — ToF. ToF performance was highly representative of the encompassing RV at-risk group (FIG.20) with notably poor RV dilation performance. As shown in FIG. 21, the RV at-risk group RV dilation performance improved when excluding ToF (but not other RV at-risk CHD lesions), highlighting the challenge in predicting RVEDV z-score >4 for ToF specifically. When using a secondary RV volume-specific model, performance in ToF remained poor for RVEDV z-score >4 (AUROC: 0.65), but was improved for higher and more clinically relevant cutoffs such as indexed RVEDV >160 mL / m2 (AUROC: 0.76) and >180 mL / m2 (AUROC 0.78) (FIG. 22).

[0248] FIG. 20 illustrates Performance of Electrocardiogram-Based Deep Learning Model in Tetralogy of Fallot. Model performance on tetralogy of Fallot (blue), RV at-risk (orange), and RV at-risk without tetralogy of Fallot (green) subgroups evaluated with receiver operating (left) and precision-recall (right) curves for: (A) left ventricular ejection fraction (LVEF) < 40%; (B) Left ventricular end-diastolic volume (LVEDV) z-score > 4; (C) right ventricular ejection fraction (RVEF) < 35%; and (D) Right ventricular end-diastolic volume (RVEDV) z-score > 4. AUROC and AUPRC metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 Note there were insufficient LVEF and LVEDV outcomes for RV at-risk without tetralogy of Fallot to generate meaningful AUROC and AUPRC curves.

[0249] Abbreviations: area under the receiver operating curve (AUROC), area under precision recall curve (AUPRC), positive predictive value (PPV)

[0250] FIG. 21 illustrates Sensitivity Analysis of RV At-Risk Model Performance to Excluding Each RV At-Risk Lesion. Forest plot showing area under the receiver operating (AUROC; red) and precision recall (AUPRC; black) curve performance for the overall cohort, the RV at-risk subgroup, and the RV at-risk cohort when excluding specific RV at-risk lesions for right ventricular end-diastolic volume (RVEDV) z-score > 4. 95% confidence intervals are shown using bootstrapping.

[0251] FIG. 22 illustrates Performance of Right Ventricle Dilation-Specific Model in Tetralogy of Fallot. Model performance of the right ventricular volume-specific model on tetralogy of Fallot (blue) evaluated with receiver operating (left) and precision-recall (right) curves for: (A) right ventricular end diastolic volume (RVEDV) z-score > 4; (B) right ventricular end diastolic volume index (RVEDVi) > 160 mL / m2; (C) right ventricular end diastolic volume index (RVEDVi) > 180 mL / m2. AUROC and AUPRC metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping.

[0252] Abbreviations: positive predictive value (PPV).

[0253] Survival analysis

[0254] In the testing cohort, median follow-up after CMR was 6.4 years (Q1-Q3: 3.2-10.7 years), with 4.4% of patients experiencing all-cause mortality at a median age of 24.1 years (Ql- Q3: 18.3-43.2 years). Given the established use of CMR to risk-stratify patients with ToF, survival after CMR was assessed in this cohort when stratified as high- or low-risk based on AI- ECG predictions of LV or RV dysfunction. There was significantly lower survival in ToF patients (FIG. 23) with LV or RV dysfunction based on AI-ECG predictions. Similar trends were identified for the overall cohort, LV at-risk patients, and RV at-risk patients (FIGS. 24A-C).

[0255] FIG. 23 Survival Analysis in Tetralogy of Fallot Based on AI-ECG Predictions

[0256] Kaplan-Meier survival curves on the Tetralogy of Fallot subgroup for patients deemed high-risk (red) vs low-risk (green) based on artificial intelligence-enhanced electrocardiogram predictions of left ventricular ejection fraction (LVEF) and right ventricular ejection fraction (RVEF) outcomes, using the Youden Index as a cutoff. Statistics based on logrank testing. Number at risk inset below plots. 95% Cis are shown in shaded regions.

[0257] FIGS. 24A-C illustrate Survival Analysis Based on AI-ECG Predictions. Kaplan- Meier survival curves on the (A) overall internal cohort, (B) left ventricle (LV) at-risk subgroup,PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 and (C) right ventricle (RV) at-risk subgroup for patients deemed high-risk (orange) versus low- risk (blue) based on AI-ECG predictions of left ventricular ejection fraction (LVEF) and right ventricular ejection fraction (RVEF) outcomes, using the Youden Index as a cutoff. Statistics based on log-rank testing. Number at risk inset below plots.

[0258] Saliency mapping

[0259] In an attempt to interpret the model, saliency mapping and median waveform analysis were performed. As shown in FIG. 25, ECGs at high risk of LV dysfunction had lower amplitude and widened QRS complexes with inverted T waves in V4 to V6. Saliency mapping demonstrated that the most influential segments were QRS complexes and T waves in V4 and V6. For LV dilation, saliency mapping demonstrated that the most influential segments were QRS complexes of lateral precordial leads, with high-risk ECGs having higher amplitude QRS complexes in these regions.

[0260] FIG. 25 Explainability of Artificial Intelligence-Enhanced Electrocardiogram Outcome Predictions

[0261] Averaged median electrocardiogram waveform from the 25 highest (red) and 25 lowest (green) predictions for LVEF, LVEDV, RVEF, and RVEDV. Saliency mapping shows more (dark blue) and less (light blue) contributory regions of the electrocardiography in the background of each lead waveform. Abbreviations as in FIGS. 15A-D.

[0262] ECGs at high-risk of RV dysfunction had widened QRS complexes in VI to V4 with inverted T waves in all precordial leads. Saliency mapping suggested the most influential segments were QRS complexes in V4 and V6, and T waves in V4. Finally, ECGs at high risk of RV dilation had similar patterns and saliency maps compared to RV dysfunction. More specifically, they had widened QRS complexes with inverted T-wave in VI to V4. In addition, saliency mapping demonstrates the most influential segments were QRS complexes in V4 and V6, and T waves in nearly all precordial leads (i.e., except V5).DISCUSSION

[0263] This work provides an exemplary embodiments of a developed and externally validated AI-ECG algorithm to predict LV and RV dysfunction and dilation using gold standard CMR metrics obtained from a heterogeneous cohort of pediatric and adult patients predominantly with CHD (FIGS. 13A-13C). The model achieved similar performance during internal testing and external validation on diverse cohorts despite different baseline characteristics, suggesting model robustness and generalizability. Performance was lesion-dependent and consistently lowest in functionally single ventricles, highlighting the inherent complexities of CHD not addressed in AI-ECG algorithms to date. Importantly, the model was shown to be capable of being refined to predict outcomes that are integral to clinical decision-making in this at-riskPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 population (FIG. 22). Patients identified at high-risk of ventricular dysfunction had lower survival in the ToF cohort, suggesting potential prognostic value. Finally, saliency mapping and median waveform analysis provided a framework to develop new insights into clinically relevant ECG characteristics predictive of LV and RV dysfunction / dilation. Altogether, these findings demonstrate the promise of AI-ECG to inexpensively screen for biventricular dysfunction / dilation in CHD, which may facilitate improved access to care and help prioritize patients for further imaging studies and / or interventions.

[0264] CMR in CHD

[0265] The indications for CMR in CHD are broad and include evaluating anatomy, physiology / hemodynamics (e.g., pulmonary / systemic flow, collateral flow evaluation), myocardial scarring, valve function, and biventricular size / function.

[0266] The known limitations of echocardiography for pediatric and adult patients to assess the RV8, makes referral for right-sided lesions a common indication for CMR. Common right-sided lesions include ToF, pulmonary atresia, atrial septal defects, and Ebstein anomaly. As these patients age, they are at increased risk of RV myopathy, which can lead to significant morbidity and mortality.

[0267] Lesion-specific guidelines provide recommended CMR surveillance frequency to help guide the timing of interventions and heart failure management. For example, in adults with ToF at risk of pulmonary regurgitation and RV or LV dilation / dysfunction, guidelines recommend CMR surveillance every 12 to 36 months. In ToF, biventricular function and size inform the need for pulmonary valve replacement, with consensus criteria for pulmonary valve replacement dependent on RVEDV, RVEF, and LVEF.

[0268] Although CMR carries these tremendous benefits specific to the pediatric cardiology population, it is limited by being time-, resource-, and cost-intensive, making it of interest to develop a convenient, standardized, and inexpensive tool to help inform timing of CMR.

[0269] AI-ECG clinical significance and implications

[0270] Compared with CMR, AI-ECG has several advantages: 1) it is rapid, conveniently obtained, and cost-effective; 2) it is standardized, without being subject to inter-rater and intrarater variability; and 3) it is safe and there are no practical limitations hindering use. In this study, AI-ECG was utilized to infer only a portion of measurements obtained from CMR (i.e., biventricular size / function). The clinical implementation of this AI-ECG algorithm may serve as a screening tool to inform timing of CMR, as well as to improve access to care.

[0271] As a screening tool, this algorithm may have economic value and lower the costs associated with care of individuals with CHD20 by reducing the frequency of noninvasivePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 imaging (e.g., echocardiography, CMR), which may even reduce the frequency of diagnostic and / or interventional catheterizations. As a thought example, the model has the capacity to achieve a negative predictive value of 99. 1% for LV dysfunction in the overall internal test cohort and 99.5% externally, with the potential to reduce CMRs for LV size / function indications by 75% to 79%. For RV dysfunction, negative predictive values of 98.9% and 96.0% were achieved, respectively, with the potential to reduce CMRs for RV size / function indications by 76% to 83%. Finally, in the case of the ToF subgroup, the secondary' RV dilation-specific model has the potential to reduce CMRs for RV size indications by 28% at a 90% sensitivity to predict RVEDV index of 160 mL / m2 (i.e., one of the proactive criteria for pulmonary valve replacement in ToF). On the other hand, the AI-ECG predictions may also help identify high-risk patients at an earlier age who will require closer monitoring and / or earlier CMR studies.

[0272] Finally , this algorithm may help improve access to care, especially in centers / areas without reliable access to CMR (and thus limited measurements of RV size / function). Currently, approximately two-thirds of the world population does not have access to specialized CHD services, with the majority of CHD patients now adults. This democratization of specialty expertise may similarly be valuable for hospitals with low pediatric volumes and / or limited pediatric cardiology experience.

[0273] Model explainability and AI-ECG insights

[0274] Model explainability increases the transparency / interpretability of models for clinicians and may aid clinicians in identifying ECG signatures resembling myopathy. High-risk features for RV dysfunction and dilation include widened QRS complexes and inverted T waves in precordial leads. These findings differ from other adult AI-ECG saliency maps of RV pathology, highlighting the unique considerations in CHD. Notably, widened QRS complexes — especially in ToF — are associated with morbidity', mortality, and RV pathology. Salient features identified for LVEF are similar to the inventors’ previous work in children without major CHD and adults with structurally normal hearts; however, unique high-risk features were identified including reduced QRS amplitude and widened QRS intervals, similarly demonstrating that the distinct ECG features in a CHD cohort require tailored AI-ECG models. Interestingly, LVEDV saliency maps and high-risk features are more similar to the inventors’ previous work.

[0275] Several insights into CHD-specific AI-ECG challenges were also gained by performing subgroup analysis. Most notably, there was poorer performance for RVEDV z-score >4 for ToF (FIG. 20), and all outcomes for functionally single ventricles (FIG. 18). In the case of ToF, the high prevalence of outcomes with frequently wide QRS duration and right bundle branch block at baseline were hypothesized to have led to subtle ECG changes that were not easily identified by the model; in contrast, at a higher RVEDV cutoff (e.g., RVEDV index >160PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 or >180 mL / m2), performance was similar to the overall model in predicting RVEDV z-score >4, possibly explained by more obvious ECG changes at this extreme. In the case of single ventricles, there can be several attributable factors for poorer performance, including the following: 1) wide range of distinct anatomic / physiologic considerations; 2) unique postoperative physiology such as septating 1 morphologic ventricle (e.g., double inlet left ventricle) into 2; and 3) relatively smaller sample sizes in this extremely heterogeneous subset of patients.

[0276] Study limitations and future directions

[0277] First, these findings are limited to patients routinely referred for CMR with a measurable secondary contributing chamber. In addition, the requirement of CMR data excludes patients with pacemakers and defibrillators. Future avenues to mitigate this limitation include obtaining biventricular function / size data from cardiac computed tomography or 3 -dimensional echocardiography. Patients with body surface area <1 m2 were excluded, thereby excluding young children. Second, a referral bias for CMR testing may lead to over-representation of certain lesions and sicker patients in the cohort, although performance was consistent across multiple care centers with different patient populations. Third, although external validation was achieved, it is of great interest to obtain external validation for each CHD lesion, and across multiple institutions globally to capture more diversity. Fourth, only 1 example of thresholding (i.e., maximal sensitivity and specificity) was used in evaluation of model performance, as further consideration (e g., weighted loss function) is required to weigh the impact of resultant false positives (which may lead to unnecessary referrals to CMR) and false negatives (which may- lead to clinical consequences of missed ventricular pathophysiology), as well as optimally set thresholds across institutions. Further multi center external validation is warranted to refine thresholds for clinical implementation. Similarly, multicenter collaboration via federated learning may help improve traimng / testing sample sizes and enhance diversity, which may further improve performance within each specific lesion (e.g., single-ventricle patients) as well as outcome (e.g., RV dilation in ToF). In addition, only ECG inputs were utilized; multimodal inputs may lead to improved model performance, especially for the complex / heterogeneous single ventricle subgroup. Until then, the model is likely best served for patients without functionally single ventricles. Fifth, discrete cutoffs corresponding to > moderate dysfunction or dilation were selected, although other cutoffs could have been considered as truth labels.Although all-cause mortality is routinely coded within the database, it is possible that positive cases are undocumented for survival analyses. Further assessment of prognostic value is needed across a range of other individual CHD lesions. Although saliency mapping provides insight into model behavior, its limitations must be noted; other methodologies such as layer-wise relevance propagation may be considered for ongoing efforts to enhance transparency and build clinicianPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 acceptance. Last, the diagnostic categories in this study are quite heterogeneous, with grouping choices performed using clinical experience.

[0278] Conclusions

[0279] Our findings demonstrate the promise of Al -ECG to inexpensively screen for and / or predict biventricular dysfunction and dilation in patients with and without CHD as defined by CMR metrics. This tool may facilitate prioritization of patients for future interventions / imaging studies, decrease costs by reducing the frequency of echocardiograms and CMRs, provide meaningful insight into novel ECG waveforms suggestive of biventricular dysfunction / dilation, and potentially reduce disparities by improving access to care. Future multicenter collaboration and prospective trials are warranted.

[0280] Example 3: Electrocardiogram-based deep learning to predict left ventricular systolic dysfunction in pediatric and adult congenital heart disease

[0281] The purpose of this example is to provide exemplary data and uses cases of a trained model described herein.

[0282] INTRODUCTION

[0283] Medical and surgical advancements have led to improved survival in children with congenital heart disease, with more than 90% of children with congenital heart disease reaching adulthood and more than 1 million adults living with congenital heart disease in the USA and Europe. This growing population remains at increased risk of heart failure, a leading cause of death in people with congenital heart disease, with higher rates in more complex forms of disease. Pharmacological interventions targeting neurohormonal pathways and device implantation play integral roles in improving heart failure symptoms and survival in the general adult population. However, there remains a paucity of evidence-based therapies specific to heart failure in congenital heart disease, making it of interest to improve preventive strategies by conveniently and inexpensively detecting early markers, such as left ventricular systolic dysfunction (LVSD). LVSD is independently associated with cardiovascular events in congenital heart disease, with guideline-directed medical therapy (GDMT) and cardiac resynchronization therapy associated with improvement in left ventricular ejection fraction (LVEF).

[0284] Artificial intelligence-enhanced electrocardiogram (AI-ECG) has shown promise as an inexpensive, ubiquitous, and non-invasive screening tool to detect LVSD in the general adult population. However, there is a paucity of AI-ECG applications to predict LVSD in pediatric cardiology, limited to patients without major congenital heart disease or select patients undergoing cardiac MRI. There remains a large unmet need to leverage AI-ECG to predict LVSD across the spectrum of pediatric congenital heart disease lesions, which have substantiallyPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 different epidemiology, anatomic structure, and ECG patterns that limit generalizability of applying adult AI-ECG algorithms. In this study, this gap was addressed by developing and externally validating an AI-ECG model on a comprehensive pediatric and adult population with congenital heart disease to predict imaging-defined LVSD.

[0285] METHODS

[0286] Internal study population and patient assignment

[0287] Our study adheres to the TRIPOD + Al guidelines. Patient data from Boston Children’s Hospital (Boston, MA, USA) up to January, 2023 were used. Inclusion criteria comprised any patient with at least one echocardiogram with a recorded LVEF. Patients with cardiomyopathy and patients without congenital heart disease were also included to enrich the training set with overlapping pathophysiology that ultimately leads to pediatric heart failure, and to broaden application to the diverse cohort encountered in the pediatric cardiology clinic.

[0288] Each qualifying echocardiogram was paired to the closest ECG, with only ECG- echocardiogram pairs 2 days or less apart included. ECGs that did not pass quality control were removed. The remaining ECG-echo- cardiogram pairs comprised the main study cohort. A group-stratified design was implemented to minimize data leakage by partitioning the main cohort at the patient level into training (70%) and test (30%) sets.

[0289] Restricting qualifying echocardiograms only to those with available LVEF would lead to selection bias, especially for specific lesions where LVEF is less often reported (hypoplastic left heart syndrome [HLHS], tricuspid atresia, and L-loop transposition of the great arteries [TGA]). To obtain LVEF in these cases, cardiac MRI is at times required. To address this limitation, model performance was assessed on ECG-cardiac MRI pairs 30 days or less apart without an intermediate intervention. Only patients outside the main echocardiogram cohort were included to properly assess this systematic difference.

[0290] External study population

[0291] For external validation, patient data from the Children’s Hospital of Philadelphia (Philadelphia, PA, USA) were obtained. Inclusion criteria comprised echocardiograms with a recorded LVEF, and at least one ECG-echocardiogram pair 2 days or less apart. Each qualifying echocardiogram event was paired with the closest ECG. ECGs that did not pass quality control were removed. The remaining ECG-echocardiogram pairs comprised the external cohort.

[0292] Data retrieval

[0293] ECG lead placement is consistent, even in the case of dextrocardia. At both Boston Children’s Hospital and the Children’s Hospital of Philadelphia, all raw ECG signals were obtained from the MUSE ECG data management system (GE Healthcare; Chicago, IL, USA).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0294] For both institutions, LVEF was obtained from echocardiogram reports, where the left ventricle always corresponds to the morphological left ventricle. At Boston Children’s Hospital, LVEF was calculated via the bullet method. At the Children’s Hospital of Philadelphia, LVEF was calculated via the biplane Simpson method.

[0295] At Boston Children’s Hospital, patient congenital heart disease lesions were identified based on the institutional Fyler coding system. At the Children’s Hospital of Philadelphia, congenital heart disease lesions were identified via ICD-9 and ICD-10 codes. At both institutions, paced patients were identified based on ECG diagnoses of dual chamber pacing or ventricular pacing.

[0296] Quality control and data preprocessing

[0297] ECGs less than 10 s long or missing lead information were discarded. Less than 2% of ECGs did not pass quality control, which was deemed to occur at random (e.g., accidentally unconnected ECG leads). The passing ECGs then were resampled to 250 Hz and underwent a high pass filter and trimming to 2048 samples (approximately 8 s) to facilitate conveniently working with convolutional neural networks. Details of quality control and preprocessing have been published previously.

[0298] Outcomes

[0299] The primary outcome was LVEF of 40% or less (quantitatively at least moderate dysfunction). Secondary outcomes included LVEF of 50% or less (quantitatively at least mild dysfunction) and LVEF of 30% or less (quantitatively severe dysfunction). As secondary analyses, time to mortality and time to LVSD onset were evaluated.

[0300] Model selection, architecture, and training

[0301] The model was developed on the training set, which, in accordance with the inventors’ previous work, was further partitioned into 95% for training and 5% for validation and hyperparameter tuning. 12-lead ECG samples of length 20148 were used as inputs to a convolutional neural network with residual block architecture, which has been adapted for unidimensional signals as previously described.

[0302] The final hyperparameters were obtained via a grid search, as follows: kernel size (3, 9, 17), batch size (8, 32, 64), and initial learning rate (0 01, 0 001, 0 0001). The average cross-entropy was minimized using the Adam optimizer. A maximum of 150 epochs were used with early stopping based on validation loss with a patience of five epochs. This choice was empirically based on the inventors’ previous work to limit computational expense and target generalizability. The model with the lowest validation loss during hyperparameter tuning was selected as the final model (kernel size 9, batch size 64, learning rate 0 001).

[0303] Performance evaluation and statistical analysesPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0304] Multiple ECG- echocardiogram pairs per patient were allowed in the training cohort. By contrast, model performance was evaluated on test groups using one randomly selected ECG-echocardiogram pair per patient. As an ancillary approach, model performance was assessed on the first or last available ECG-echocardiogram pair.

[0305] Given the imbalanced dataset, both the area under the receiver operating curve (AUROC) and area under the precision-recall (i. e. , positive predictive value [PPV]- sensitivity) curve (AUPRC) were computed. Other performance metrics evaluated included PPV, negative predictive value (NPV), sensitivity, and specificity. These metrics were calculated based on a low-risk threshold (achieving 95% sensitivity in the training set) and high-risk threshold (achieving 95% specificity in the training set). Confidence intervals were obtained via resampling with 1000 bootstraps.

[0306] Subgroup analyses

[0307] Subgroup analyses were done across a range of congenital heart disease lesions and age subgroups on test sets using all available ECG-echocardiogram pairs 2 days or less apart.

[0308] Benchmarking model performance

[0309] To benchmark the model, the performance of the current model was compared with a previous AI-ECG model for children without major congenital heart disease. The benchmarking test cohort comprised all known patients with congenital heart disease who were independent from both model training sets. A subgroup analysis was done to benchmark across individual congenital heart disease lesions. Model performance was compared using the DeLong test.

[0310] Time-to-event analysis

[0311] Time-to-event analysis was done for LVSD onset and all-cause mortality. Time- to-LVSD onset analysis (where the event of interest was LVEF <40%) was done by: including only patients with multiple ECG-echocardiogram pairs and a first echo LVEF >40%; stratifying patients into three groups based on the AI-ECG classification of the first ECG-echocardiogram pair (low risk [AI-ECG probability of at least low-risk cutoff], intermediate risk [low-risk cutoff less than AI-ECG probability of at least high-risk cutoff], and high risk [AI-ECG cutoff greater than high-risk cutoff]); and assessing time-to-event after ECG within each group. Patients who did not have LVSD onset were censored at the time of last echocardiogram.

[0312] Time-to-mortality was done by stratifying patients into the same groups based on a single random ECG per patient and assessing time-to-event after ECG within each group.Patients who did not have all-cause mortality were censored at time last known alive. Patients with unknown follow-up time after ECG were excluded from survival analysis.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0313] Cox proportional hazards regression was used to evaluate AI-ECG classification association with time from ECG until the event of interest (LVSD onset or all-cause mortality). Hazard ratios (HRs) were adjusted for age, with the low-risk group as reference. Statistical comparison between groups was based on log-rank testing.

[0314] Coding language

[0315] The convolutional neural network used the Keras framework with a Tensorflow (Google) backend using Python 3.9. Deep learning was executed on institutional graphics processing units. All other pre-processing and post-processing code was written in Python 3.9 and R 4.0, which was executed locally.

[0316] Model explainability

[0317] Model interpretability was explored via median waveform analysis and saliency mapping, as described previously. The number of samples used to generate representative median waveforms and saliency maps (n=100 for the overall test cohort and n=25 for individual congenital heart disease lesions) was selected empirically based on the inventors’ previous work.

[0318] Role of the funding source

[0319] The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report.

[0320] RESULTS

[0321] The training cohort comprised 124 265 ECG-echo- cardiogram pairs (49 158 patients; median age 10-5 years [IQR 3-5—16-8]; 22 835 [46 5%] of 49 158 patients were female and 26 311 [53-5%] were male; TABLE 14). The median echocardiogram LVEF was 62- 0% (IQR 57-4—66-0), where 3381 (2-7%) of 124 265 ECG-echocardiogram pairs had an LVEF of 40% or less. The most common lesions included tetralogy of Fallot (1911 [3-9%] of 49 158 patients), cardiomyopathy (2428 [4- 9%] patients), atrial septal defects (3516 [7-2%] patients), coarctation of the aorta (2406 [4-9%] patients), and ventricular septal defects (5178 [10-5%] patients; TABLE 14). Complex lesions with lower prevalence included HLHS (450 [0- 9%] patients), L-loop TGA (251 [0- 5%] patients), and tricuspid atresia (242 [0-5%] patients; TABLE 14). 465 (0-9%) patients were ventricularly paced. 1478 (3 0%) patients died. Similar lesion and outcome breakdowns were found in the internal testing cohort, which comprised of 54 230 ECG- echocardiogram pairs (21 068 patients; median age 10-9 years [IQR 3 7-17 0]; 9813 [46-6%] of 21 068 patients were female and 11 251 [53-4%] were male; TABLE 14). 16 930 (24- 1%) of 70 226 patients had known congenital heart disease in the overall internal cohort.

[0322] The external validation cohort comprised 76 400 ECG- echocardiogram pairs (42 984 patients; median age 10- 8 years [IQR 4-9-15 0]; 19 163 [44-6%] of 42 984 patients were female and 23 815 [55 -4%] were male; TABLE 14). The median echo LVEF was slightly higherPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 at 64 -2% (IQR 60-4-67 -2) in the external validation cohort than in the training and internal testing cohorts, with a lower prevalence of outcomes (1313 [1 -7%] of 76 400 ECG- echocardiogram pairs with LVEF <40%; TABLE 14). In general, there was a lower prevalence of each lesion and ventricularly paced patients in the external cohort compared with the internal cohort.

[0323] TABLE 14: Baseline Characteristics (Training and Internal Testing data obtained from Boston Children’s Hospital, External validation data obtained from Children’s Hospital ofPhiladelphiaPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0324] AI-ECG achieved high performance to discriminate mild (LVEF <50%), moderate (LVEF <40%), and severe dysfunction (LVEF <30%) during internal (AUROC 0 95, AUPRC 0-33) and external (AUROC 0-96, AUPRC 0-25) testing (FIG. 26). The model is well calibrated for each outcome of interest. FIG. 26 illustrates Internal testing and external validation of the AI- ECG model to predict left ventricular systolic dysfunction. Performance of the AI-ECG algorithm evaluated in the internal (left) and external (right) cohorts using receiver operating (AUROC) and precision-recall (AUPRC) curves for the following outcomes: LVEF <50%, LVEF <40%, and LVEF <30%. Model performance was assessed on a single ECG- echocardiogram pair per patient: random (blue), first (red), and last (green). AUROC and AUPRC metric values for each model and outcome are inset; values in parentheses are 95% Cis. Dotted line represents chance. Shaded areas represent 95% Cis, which were generated using bootstrapping. AI-ECG=artificial intelligence-enhanced electrocardiogram. AUPRC=area under the precision-recall curve. AU ROC -area under the receiver operating characteristic curve. LVEF=left ventricular ejection fraction. PPV=positive predictive value.

[0325] Regarding performance metrics for low-risk and high-risk cutoffs, at the low-risk cutoff for LVEF of 40% or lower, sensitivity is around 90% across institutions with NPV at least 99-8% and approximately 90% of ECGs were predicted negative. At the high-risk cutoff for LVEF of 40% or lower, specificity was 97-98% across institutions with NPV at least 99 5% and PPV 13-20%.

[0326] In general, AI-ECG model performance was lower for more complex lesions and for lower prevalence lesions, with the lowest performance in dextrocardia. Modest performance was achieved across institutions for L-loop TGA and functionally single ventricle lesions such as HLHS and tricuspid atresia (FIG. 27A). Given that it is more difficult to ascertain LVEF on echocardiogram for these lesions, model performance on ECG-cardiac MRI pairs was also assessed. In this set of patients independent from the main cohort, model performance was again modest for HLHS, L-loop TGA, and tricuspid atresia (FIG. 27B). AI-ECG performance remained high across all age groups, with the highest discrimination achieved for same-day ECG-echocardiogram pairs. During benchmarking, this model outperformed a previously established model to predict LVSD in patients with any known congenital heart disease (p<0 0001; data not provided), as well as across a range of individual congenital heart disease lesions (data not provided). FIGS. 27A-B illustrate: Model performance across congenital heart disease lesion subgroups. Internal testing (blue) and external validation (red) AUROC (left column) and AUPRC (right column) performance when stratifying by a range of congenital heartPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 disease lesions (listed from highest to lowest prevalence) for the outcome of LVEF <40%. 95% Cis are shown using bootstrapping and indicated by error bars.

[0327] Model performance (AUROC and AUPRC) to predict LVEF <40% using ECG- echocardiogram pairs (red) vs ECG-cardiac MRI pairs (blue) for select lesions. Error bars indicate 95% Cis. AUPRC=area under the precision-recall curve. AUROC=area under the receiver operating characteristic curve. ECG=electrocardiogram. HLHS=hypoplastic left heart syndrome. LVEF=left ventricular ejection fraction. TGA=transposition of the great arteries.

[0328] The prognostic value of the initial ECG-echocardiogram pair to predict future LVSD was assessed (FIGS. 28A-B). Relative to low-risk patients, high-risk patients were more likely to have a future LVEF of 40% or less in the overall cohort (HR 12- 1 [95% CI 8-4-17-3]; p<0-0001), as well as in cardiomyopathy (5-8 [95% CI 3-4-10 0]; p<0- 0001) and tetralogy of Fallot subgroups (8-2 [95% CI 1 -7-39-8]; p=0 0094). AI-ECG predictions were also predictive of future all-cause mortality. Using a single random ECG per patient, patients with ECGs deemed high-risk were more likely to have all-cause mortality compared with those deemed low-risk. Similar prognostic trends were noted in patients independent from the training cohort who presented to the cardiology clinic and did not have an echocardiogram within 2 days of the visit (data not provided).

[0329] FIGS. 28A-B illustrate Future left ventricular systolic dysfunction or mortality based on AI-ECG classification (A) Incidence of future LVEF <40% for the overall test cohort (left), patients with cardiomyopathy (middle), and patients with tetralogy of Fallot (right) initially with LVEF >40%, stratified by initial network classification (low-risk in green, intermediate-risk in blue, and high-risk in red). Number of patients at risk over the 10-year period given below the graphs. The shaded areas surrounding the curves represent 95% Cis. (B) Survival analysis when stratifying patients as low-risk, intermediate- risk, or high-risk based on AI-ECG left ventricular systolic dysfunction probabilities for the overall cohort (left), cohort with cardiomyopathy (middle), or cohort with tetralogy of Fallot (right). Color-coded HR with 95% CI is given below the graphs (obtained via Cox regression analysis when adjusting for age, with the low-risk group as a reference). The shaded areas surrounding the curves represent 95% Cis. AI-ECG=artificial intelligence-enhanced electrocardiogram. HR=hazard ratio. LVEF=left ventricular ejection fraction.

[0330] Saliency mapping and median waveform analysis were done to generate hypotheses of ECG features driving LVSD model predictions (FIG. 29). In the overall test cohort, salient features included the QRS complexes for V2 and V5-V6, as well as the V6 T wave. Predicted high-risk features of LVSD include deep S waves in V2 and tall R waves in V5- V6, with inverted T waves in the lateral precordial leads. In cardiomyopathy, nearly identicalPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 saliency and high-risk features were identified. In HLHS, similar saliency and high-risk features were identified, with the exception of lower amplitude high-risk R waves in V5-V6. Tricuspid atresia had similar high-risk features to HLHS, with QRS complexes being salient in precordial leads V2-V6. For tetralogy of Fallot and L-loop TGA, QRS complexes across most precordial leads were salient. Deep S waves in V2 were also high-risk for tetralogy of Fallot and L-loop TGA, with other high-risk features including inverted T waves in lateral precordial leads and wide QRS complexes.

[0331] FIG. 29 illustrates explainability of AI-ECG predictions. Median waveforms generated in each lead using ECGs from the highest (red) and lowest (green) AI-ECG predictions of the overall test cohort, as well as cardiomyopathy, tetralogy of Fallot, HLHS, tricuspid atresia, and L-loop TGA subgroups. Saliency mapping demarcates regions of the ECG waveform having greatest (dark blue) and least (light blue) effect on each outcome. Saliency was averaged over the highest predicted ECGs for left ventricular ejection fraction <40%. AI-ECG=artificial intelligence-enhanced electrocardiogram. HLHS=hypoplastic left heart syndrome. TGA=transposition of the great arteries.

[0332] Given the unique considerations and adverse impact of ventricular pacing in congenital heart disease, model performance and prognostic value in this specific subgroup were also assessed. Modest model performance was achieved across internal and external cohorts (data not provided). Compared with low-risk paced patients, high-risk patients had an increased risk of all-cause mortality (p=0 018). Salient features similarly included QRS complexes in precordial leads (V2-V4, V6), as well as V6 T waves. High-risk features included deep and wide QRS complexes in V2-V3 and inverted T waves in V6.

[0333] DISCUSSION

[0334] This study represents a comprehensive application of ECG-based deep learning to heterogeneous pediatric and adult congenital heart disease to predict LVSD. After training on more than 100,000 ECG-echocardiogram pairs from around 50,000 patients, model performance was high for a range of congenital heart disease lesions across two large children’s hospital health-care systems. AI-ECG provided prognostic value, predicting future LVSD and all-cause mortality in populations with congenital heart disease. This study shows the promise of AI-ECG to inexpensively screen for or predict LVSD in pediatric and adult congenital heart disease, which might facilitate improved access to care, help prioritize patients for further studies or interventions, and inform ventricular pacing strategies.

[0335] Heart failure accounts for approximately 20% of hospital admissions for adults with congenital heart disease and is a substantial (20-40%) cause of all-cause mortality in adults with congenital heart disease. Factors contributing to myocardial dysfunction and heart failure inPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 congenital heart disease are complex and include a multitude of hemodynamic and electrophysiological derangements (e.g., volume or pressure loading, sequelae of previous surgeries, arrhythmia, chronic ventricular pacing, and ventricular scarring or fibrosis). Although recent data suggest sodium-glucose cotransporter-2 inhibitors reduce heart failure hospitalization rates in adults with congenital heart disease, there remains limited data showing GDMT for heart failure in adults is similarly effective in children or adults with congenital heart disease.

[0336] The prospect of developing novel approaches for inexpensive and convenient early screening for closely linked markers of heart failure, such as LVSD, motivated efforts to develop an AI-ECG tool to predict LVSD across a range of congenital heart disease lesions. Prediction of LVSD in congenital heart disease is advantageous given that it provides a snapshot of ventricular health, carries prognostic value, and has the potential to be modified with GDMT and cardiac resynchronization therapy. Additionally, earlier identification of patients with congenital heart disease at risk for LVSD — in particular, more severe degrees of dysfunction — might help tailor treatment strategies and intensities for this group. This tool has potential for multiple clinical applications including screening for LVSD in congenital heart disease (which might reduce health-care costs associated with unnecessary echocardiograms), identifying patients at greater risk of LVSD (which might lead to closer monitoring or earlier initiation of GDMT), and conveniently tracking risk of LVSD across the lifespan.

[0337] As a screening tool, the low-risk cutoff achieved around 90% sensitivity across institutions, with the potential to decrease obtaining around 60% of echocardiograms (for detecting mild dysfunction) to around 90% of echocardiograms (for detecting moderate or severe dysfunction) at NPVs of at least 99-5%. Low-risk patients also had significantly lower future LVSD and all-cause mortality, which might facilitate reduced follow-up frequency. By contrast, the high-risk cutoff was highly specific (around 95%) across institutions, with a PPV of 13-20% to detect LVEF of 40% or less. Furthermore, high-risk patients had significantly higher future LVSD and all-cause mortality, which might trigger more thorough follow-up and potentially earlier initiation of GDMT. Finally, for intermediate-risk patients, the standard of care can be continued.

[0338] The model is well calibrated across all risk groups and could theoretically serve as a tracking tool across the lifespan to provide convenient assessment of ventricular health and a response to interventions shown to improve ventricular function (e.g., GDMT in congenital heart disease). This tool could be especially valuable for low-resource settings and children with stage B heart failure (given current recommendations for echocardiograms approximately every 6 months).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0339] Ventricular pacing in congenital heart disease — and more specifically in patients with functionally single ventricles — is associated with increased risk of heart transplant and mortality. Additionally, cardiac resynchronization therapy is associated with improvement in left ventricular systolic function. Taken together with the distinct ECG waveforms of ventricular pacing, it was of interest to investigate this subgroup separately. Similar to the cohort with congenital heart disease, this early case study on patients with ventricular pacing might inspire clinical translation to screen for LVSD (which might reduce health-care echocardiogram use), predict future LVSD or mortality (which might help inform on the highest-risk candidates for pacemakers), and identify salient and high-risk features of LVSD (which might inform lead placement to optimize ECG waveforms beyond using narrow QRS duration).

[0340] Congenital heart disease-specific models are desirable to account for congenital heart disease-specific anatomical structures and ECG patterns. Similar reasoning might explain why adult AI-ECG algorithms have limited generalizability to pediatric populations. Model explainability might provide insight into the underlying pathology that can subtly manifest within ECGs for patients with congenital heart disease. For example, common salient features across lesions include V2 QRS complexes, with high-risk features of deep S waves. These findings were similarly highlighted by Sangha and colleagues in a deep learning-based ECG model for the general adult population. Mechanistically , one could speculate that the model might be focusing electrocardiographically on the anteroseptal left ventricle, recognizing delayed myocardial activation. This interpretation is further reinforced by similar patterns noted in patients with pacemakers, suggesting a mechanism independent of the native conduction system. Delayed myocardial activation can lead to cardiac dyssynchrony, which is linked to LVSD and heart failure. Future work is needed to rigorously investigate this hypothesis in relation to underlying pathophysiology.

[0341] There are several limitations of this work. First, inclusion criteria consisted of echocardiograms with recorded LVEF, which inserts selection bias for certain lesions that are less likely to have echocardiograms with recorded LVEF (e.g., HLHS, L-loop TGA, and tricuspid atresia). This limitation was addressed by effective model performance using ECG- cardiac MRI pairs in patients outside the main cohort. Selection bias also exists when training on contemporaneous ECG-echocardiogram pairs, although reassuringly the model still holds predictive value in patients presenting to the clinic without an echocardiogram. Second, although AUROC and AUPRC performance remain high between internal and external cohorts (showing generalizability), performance was lower for more complex and less prevalent lesions. To address this limitation, future studies will aim to include multicenter collaboration via federated learning to compile a larger set of heterogeneous data across institutions, incorporation of clinicalPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 variables, and use of various training techniques (e.g., ensemble models, data augmentation, and regularization). Additionally, although performance was high in simple lesions across institutions, performance was only evaluated in large referral centers, thus warranting future multicenter external validation across the spectrum of care levels (e.g., secondary, tertiary , and quaternar ). Third, model inputs require access to digital waveform data, which impedes translation to low-resource settings where such data are unavailable. Future efforts to address this limitation include training a model using ECG image inputs. The pacemaker cohort provided proof-of-concept for AI-ECG applications, which requires further comprehensive investigation. Pragmatic randomized clinical trials to prospectively determine effectiveness of AI-ECG as an ECG screening tool, to guide clinical implementation, and inform cost-effectiveness are warranted. Although heart failure is a leading cause of death in adults with congenital heart disease, all-cause mortality rather than cardiac mortality was used for a survival analysis. Although saliency mapping provides insight into model behavior, limitations must be noted. Although the model does show prognostic value, other approaches to directly predict future heart failure or mortality should be considered. Finally, patient subgroups were identified via the available coding infrastructures within each institution, which are limited by potential miscoding errors.

[0342] In conclusion, these findings show the promise of AI-ECG to inexpensively screen for and predict future LVSD in pediatric and adult congenital heart disease. This tool might facilitate improved access to care, help prioritize patients for future interventions or studies, and potentially inform ventricular pacing strategies. Future multicenter collaboration and pragmatic randomized clinical trials are warranted.Example 4: Expert-Level Diagnosis of Pediatric 12-Lead ECG

[0343] The purpose of this example is to provide exemplary data that may be used for training a model described herein for diagnosis of cardiac abnormalities using pediatric 12-lead ECGs.Study Population and Patient Assignment

[0344] Patient data was utilized from Boston Children’s Hospital. Inclusion criteria consisted of any patient with at least one ECG between 2000-2022 without missing metadata (e.g., missing medical record number, ECG event number, reading provider, or ECG diagnosis). Note that pediatric and adult congenital patients were included given: 1) adult congenital patients have unique ECG pattems / findings that are related to their underlying congenital heart lesion and their sequalae of cardiac interventions; 2) pediatric cardiologists are frequently caring for thesePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 patients and interpreting their ECGs; 3) there is a global shortage of the adult congenital heart disease workforce.

[0345] To optimize training label accuracy, ECGs were removed from less experienced readers (i.e., providers with < 5,000 ECGs). ECGs failing to pass quality control (for quality control details, see “Quality Control and Data Preprocessing” below) were also removed. Finally, ECGs with equivocal (e.g., “probably normal ECG variant”) or outdated (e.g., “counterclockwise rotation”) diagnoses were removed. The remaining ECGs comprised the main cohort, which was then partitioned at the patient level into training (75%) and testing (25%) cohorts.Data Retrieval

[0346] ECG waveforms (I, II, and V1-V6) were obtained from the MUSE ECG data management system, with each lead corresponding to a one-dimensional vector sampled at 250 Hz for a 10 second duration (2500 samples). Leads III, aVF, aVL, and aVR were reconstructed using Einthoven’s law and the Goldberger equation.

[0347] ECG diagnoses, age, sex, and congenital heart lesion diagnoses were identified based on the institutional Fyler coding system. During the study period, all ECGs were read using a custom internal software (ECG Reader) which requires physicians to select one or more pre-determined ECG diagnostic codes, providing a cleanly labeled dataset for training and testing purposes.Quality Control and Data Preprocessing

[0348] Briefly, ECGs without 2500 samples or missing lead information were removed.For each passing ECG, a high pass filter and trimming process was then applied to 2048 samples (~8 seconds) for ease of working with convolutional neural networks.Definition of Primary and Secondary Outcomes

[0349] The primary composite outcome was ECG diagnosis of any abnormality (defined as any diagnostic code other than “Normal for age” and “Sinus arrhythmia”). Secondary outcomes included individual ECG diagnoses that are especially pertinent to pediatric ECG screening — namely WPW and prolonged QTc. As secondary analyses, time-to-diagnosis of ECG abnormality, WPW, and prolonged QTc (see “Time-to-Event Analysis” below) were also evaluated.

[0350] Model Selection, Architecture, and Training

[0351] The convolutional neural network involves a residual block architecture that has been adapted for unidimensional signals. Briefly, the model was developed exclusively on the training set, of which 5% was designated for validation and hyperparameter tuning. The input to the convolutional neural network was 12 x 2048 ECG samples. The final hyperparameters were obtained via a grid search: kernel size [3, 9, 17], batch size [8, 32, 64], and initial learning ratePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 [0.01, 0.001, 0.0001], The average cross-entropy was minimized using the Adam optimizer. Maximum 150 epochs were used with early stopping based on validation loss. Final hyperparameters for this model were kernel size 17, batch size 32, learning rate 0.001.

[0352] Given the known role of age and sex on ECG characteristics, a second model was created that incorporated age / sex as inputs along with ECG waveforms (AI-ECG + age + sex). For the AI-ECG + age + sex model, final hyperparameters were kernel size 9, batch size 32, learning rate 0.001.Performance Evaluation and Statistical Analyses

[0353] Model performance was evaluated exclusively on the testing cohort. Given the imbalanced prevalence for select ECG diagnoses (e.g., WPW and prolonged QTc), both area under the receiver operating curve (AUROC) and area under the precision-recall (i.e., positive predictive value (PPV)-sensitivity) curve (AUPRC) were computed. Model performance was benchmarked to commercial (GE MUSE) ECG interpretations. Other performance metrics assessed include PPV, negative predictive value (NPV), sensitivity, specificity, Fl, and accuracy. Given the objective to obtain highest overall accuracy in an imbalanced dataset, these metrics were calculated based on thresholds achieving the optimal Fl score. The Fl score is defined as the harmonic mean of the precision and sensitivity, symmetrically representing both within one metric. Fl scores range from 0 (no precision and sensitivity) to 1 (perfect precision and sensitivity). 95% Confidence intervals were obtained via resampling with 1,000 bootstraps.Subgroup Analyses

[0354] Age and sex are known to impact ECG characteristics in a healthy pediatric population. Therefore, subgroup analyses were performed across sexes and a range of ages. Within each subgroup, AUROC and AUPRC were calculated.

[0355] Time-to-Event Analysis

[0356] Time-to-event analysis was performed for the following outcomes: diagnosis of any abnormality, WPW, and prolonged QTc. For each, time-to-diagnosis onset analysis was performed by: 1) including only patients with multiple ECGs; 2) stratifying patients into two groups based on the AI-ECG classification of the first ECGs: true negative or false positive; 3) assessing time-to-diagnosis after ECG within each group.

[0357] Cox proportional hazards regression was used to evaluate AI-ECG classification association with time from ECG until the diagnosis of interest. Hazard ratios were adjusted for age and sex. Statistical comparison between groups were based on log-rank testing. Patients who did not experience ECG diagnosis onset were censored at the time of last ECG.

[0358] Readjudication and Expert AgreementPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0359] Readjudication was performed on ECGs with discrepancies between the diagnostic classification assigned by the original reader and the AI-ECG classification (using the same cutoff as above). More specifically, for each outcome (any abnormality, Wolff Parker White syndrome (WPW), prolonged QTc), 50 false positives (i.e., deemed positive by AI-ECG, but negative by the original reader) and 50 false negatives (i.e., deemed negative by AI-ECG, but positive by the original reader) were re-read by four senior pediatric electrophysiologists. These experts were blinded to the diagnoses by AI-ECG and the baseline reader.

[0360] Agreement was assessed between the baseline reader and experts, AI-ECG and experts and all four experts using Cohen’s K for agreement between two entities) or Fleiss K for agreement between more than two entities).12 Values < 0 indicate no agreement, with 0-0.20 as slight, 0.21-0.40 as fair, 0.41-0.60 as moderate, 0.61-0.80 as substantial, and 0.81-1 as near perfect agreement.

[0361] Model Explainability

[0362] Model behavior was investigated via median waveform analysis and saliency mapping. Briefly, median waveforms provide visual representations of high- and low-risk ECGs. The 100 highest predicted ECGs for WPW and prolonged QTc were used to create high-risk median waveforms. To contrast to normal ECGs, the 100 lowest predicted ECGs for any abnormality were used to create low-risk median waveforms.

[0363] Saliency mapping provides insight into important ECG patterns that contribute to model prediction. Using a Shapley Additive Explanations (SHAP) framework, saliency maps highlight ECG regions where a change in ECG voltage input corresponds to a change in output prediction. The 100 highest predicted ECGs for each diagnosis were used to create saliency maps.Software

[0364] Programming code used to perform the analyses are available upon reasonable request. The convolutional neural network used the Keras framework with a Tensorflow (Google) backend using Python 3.9. Deep learning was executed on institutional graphics processing units. All other pre- and post-processing code was written in Python 3.9 and R 4.0, which was executed locally.RESULTSPatient Population Baseline Characteristics

[0365] There were 734,800 ECGs (238,072 patients) between 2000-2022; after removing ECGs with missing metadata (18,539 ECGs), less experienced readers (77,363 ECGs), failed quality control (10,683 ECGs), and equivocal diagnoses (45,081 ECGs), there were 583,134 ECGs (201,620 patients) comprising the main cohort.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0366] The training and testing cohorts comprised of 437,350 ECGs (151,215 patients; 49% male; median age 11.6 [IQR 3.1-16.9] years) and 145,784 ECGs (50,405 patients; 49% male; median age 12.0 [3.3-17.0] years), respectively. In both cohorts, 33% had more than one diagnosis; 56% had any abnormality, 1.0% had WPW, and 5.3% had prolonged QTc (TABLE 15). The prevalence of congenital heart disease lesions and diagnoses of other ECG findings are highlighted in TABLE 15.

[0367] TABLE 15: Baseline Characteristics of Training and Testing CohortsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30 2025

[0368] Abbreviations: Congenital heart disease (CHD); atrial septal defect (ASD); complete atrioventricular canal defect (CAVC); coarctation of the aorta (CoA); double outlet right ventricle (DORY); transposition of the great arteries (TGA); hypoplastic left heart syndrome (HLHS); total anomalous pulmonary venous connection (TAPVC); ventricular septal defect (VSD); tetralogy of Fallot (ToF); electrocardiogram (ECG); non-specific ST / T wave changes (NSSTT); supraventricular tachycardia (SVT); Wolff Parkinson White syndrome (WPW); right ventricular hypertrophy (RVH); left ventricular hypertrophy (LVH); complete right bundle branch block (CRBBB).Model Performance

[0369] FIGS 30A-B illustrates AI-ECG model performance to detect ECG abnormalities. FIG. 30A illustrates performance of the artificial intelligence-enhanced ECG model alone (AI- ECG; blue) and with demographic data (AI-ECG + age + sex; orange) evaluated in the testing cohorts using receiver operating (AUROC; left) and precision recall (AUPRC; right) curves for the following outcomes: any abnormality, Wolff Parkinson White syndrome (WPW), andPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 prolonged QTc. Model benchmarked to commercial MUSE GE interpretations (grey dot).AUROC and AUPRC metric values for each model and outcome are inset below with 95% confidence intervals in brackets. Dotted line represents chance. FIG. 30B illustrates incidence of future abnormal ECG (top), WPW (middle), or prolonged QTc (bottom) for the test cohort stratified by initial network classification (true negative [TN] in green, false positive [FP] in red). Number of patients at risk over the 1-year period inset below. Hazard ratio and log-rank p-value inset.

[0370] Performance of the AI-ECG and AI-ECG + age + sex models to detect any abnormality, WPW, and prolonged QTc is shown in FIG. 30A. Excellent performance was achieved for AI-ECG to detect any abnormality (AUROC 0.94, AUPRC 0.96), WPW (AUROC 0.99, AUPRC 0.88), and prolonged QTc (AUROC 0.96, AUPRC 0.63). Model performance was nearly identical when adding age and sex as inputs (FIGS. 30A-30B), outperforming commercial MUSE interpretations for any abnormality (sensitivity 92%, specificity 48%, PPV 69%) with more pronounced differences for WPW (sensitivi ty 5%, specificity 99%, PPV 55%) and prolonged QTc (sensitivity 44%, specificity 93%, PPV 26%) (FIG. 30A).

[0371] Patients with initial false positive AI-ECG classification were more likely to have a future abnormal ECG (hazard ratio 2.0 [95% CI, 1.8-2.2]; p<0.001), WPW (HR 88 [95% CI, 35-218]; pO.OOl), and prolonged QTc (HR 3.4 [95% CI 2.9-4.0]; p<0.001) (FIG. 30B). Subgroup Analysis

[0372] FIG. 31 illustrates model performance across age and sex subgroups. Forest plots showing area under the receiver operating (AUROC; red) and precision recall (AUPRC; black) curve performance when stratifying by age, sex, and reader for diagnosing any abnormality (left), Wolff Parkinson White syndrome (WPW; middle) and prolonged QTc (right). Abbreviations: week (w); month (mo), year (y).

[0373] In a subgroup analysis (FIG. 31), there was variation in performance by age, sex, and reading provider. In general, there was lower performance for ages < 3 year old, most notably for age < 1 week for any abnormality' and prolonged QTc. Performance for ages > 18 years old were similar to the overall cohort. Performance did not vary by sex. Finally, there was slightly higher performance for a pediatric electrophysiology' specialist, most notably for prolonged QTc. There was no appreciable difference in performance when testing on a provider’s first 5,000 ECG reads (any abnormality: AUROC 0.93, AUPRC 0.96; WPW: AUROC 0.98, AUPRC 0.88; prolonged QTc: AUROC 0.96, AUPRC 0.61) to their last 5,000 ECG reads (any abnormality: AUROC 0.94, AUPRC 0.96; WPW: AUROC 0.99, AUPRC 0.87; prolonged QTc: AUROC 0.96, AUPRC 0.62).ReadjudicationPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0374] Readjudication was performed by four blinded expert electrophysiologists on ECGs with discrepancies between the baseline reader and the AI-ECG classification. As shown in FIG. 32A heatmap, AI-ECG, but not the baseline reader, clustered with the experts.Interestingly, the inter-expert Fleiss K agreement was modest across each outcome of interest, ranging from 0.21-0.49 (all p-values not significant). The expert readers on average were more likely to agree with AI-ECG than the baseline reader (FIG. 32B).

[0375] There was a 2-2 tie among experts for 20% of the any abnormality readjudication ECGs, 17% of the WPW readjudication ECGs, and 19% of the prolonged QTc readjudication ECGs. Readjudication results are similar when excluding the ECGs without majority vote.Model Explainability

[0376] FIG. 33 illustrates model explainability through visualization of high-risk (red) and low-risk (green) median waveforms for WPW (top) and prolonged QTc (bottom). Saliency mapping demarcates ECG regions with greatest (dark blue) and least (light blue) influence on each diagnosis. Model behavior analysis was performed to compare salient features noted by AI- ECG in comparison to conventional rule-based approaches for diagnostics (FIG. 33). For WPW, the most salient signals are within P waves, QRS complexes, and PR signals (limb leads I-II and precordial leads VI, V5-V6). High-risk waveforms unsurprisingly demonstrate pre-excitation (a "Delta wave”) with short PR intervals. For prolonged QTc, the most salient features were QRS complexes and T waves (precordial leads VI and V6). High-risk waveforms unsurprisingly demonstrate a prolonged QTc interval.

[0377] After training on >400,000 ECGs from nearly 150,000 patients with and without cardiac abnormalities, model performance was demonstrated to be excellent (AUROC > 0.9) for outperforming commercially available software. Model performance remained robust across a range of subgroups, and readjudication of misclassified ECGs demonstrated that four blinded senior electrophysiologists were more likely to agree with AI-ECG diagnoses than an experienced reader in these boundary cases. Finally, saliency mapping findings align with conventional rule-based methodologies implemented by humans, promoting clinician trust. Altogether, these findings demonstrate the promise of AI-ECG to assist clinicians with the rapid and reliable interpretation of ECGs, which may: 1) reduce missed diagnoses by less experienced clinicians; 2) promote screening programs; 3) facilitate improved access to expert care; and 4) decrease physician workload, which may help reduce burnout.

[0378] Several insights were developed from model performance and behavior analysis. First, AI-ECG was found to perform similarly to AI-ECG + age + sex (FIG. 30A) for primary / secondary outcomes and sinus tachycardia / brady cardia, suggesting the model is learning this demographic data that is known to relate to changes in ECG axis and intervals. Second, false-Tl-PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 positive predictions by Al -ECG were observed more likely to be positive on ECGs soon thereafter (with a hazard ratio of 88 for WPW), suggesting either initial misdiagnoses by the reader or AI-ECG detecting subtle findings that become more pronounced on follow-up ECGs (e.g., intermittent pre-excitation; borderline prolonged QTc). This is supported by the low interexpert K achieved for each outcome (FIG. 32A), which suggests discrepant reads between the baseline reader and AI-ECG are borderline or difficult boundary cases with limited agreement even by world expert readers. Finally, lower performance of any abnormality and prolonged QTc was noted in the newborn (<1 week).Example 5: Expert-Level Diagnosis of Critical Congenital Heart Disease from a Pediatric 12- Lead ECG

[0379] The purpose of this example is to provide exemplary data that may be used for training a model described herein for diagnosis of critical congenital heart disease (CCHD).INTRODUCTION

[0380] Congenital heart disease affects nearly 1% of livebirths, with -25% considered critical congenital heart disease (CCHD). Infants with late CCHD detection are at risk for cardiovascular collapse and death, underscoring the need for accurate prenatal and / or early life detection. Fetal ultrasound and pulse oximetry screening have high specificity yet moderate sensitivity to detect CCHD (especially for left-sided lesions such as hypoplastic left heart syndrome [HLHS] or coarctation of the aorta [CoA]), highlighting the need for technological advancements in CCHD screening.

[0381] Electrocardiogram (ECG) is an easily accessible, ubiquitous test frequently performed on infants <1 years old. While some recent work has demonstrated ECG-based deep learning integrated with human concepts predicting CHD, there has yet to be an infant, CCHD- specific model. In this study, this technological gap was addressed by developing and externally validated an artificial intelligence-enhanced electrocardiogram (AI-ECG) tool.METHODS

[0382] As discussed herein, “critical congenital heart disease” may be understood as a composite of the following lesions: complete atrioventricular canal defect, coarctation of the aorta, double outlet right ventricle, D-loop transposition of the great arteries, pulmonary' atresia, total anomalous pulmonary venous connections, tetralogy of Fallot, double inlet left ventricle, double inlet right ventricle, double outlet left ventricle, Ebstein anomaly, hypoplastic left heart syndrome, interrupted aortic arch, tricuspid atresia, truncus arteriosus, critical pulmonary stenosis, and critical aortic stenosis.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0383] Two convolutional neural networks were trained / tested on the first available ECG for infants <1 years old to predict CCHD and each individual CCHD lesion. The first model receives 12-lead ECG inputs as previously described; the second model receives 9-lead ECG inputs (limb leads + VI, V3, V5) to match external cohort input data structures.

[0384] The main cohort was randomly partitioned at the patient level into training (70%) and internal testing (30%) cohorts. Internal lesions were identified based on institutional Fyler codes. The external cohort included publicly available 9-lead ECG data from Guangdong Provincial People’s Hospital with labels for CCHD and select lesions.

[0385] Data retrieval, quality control, preprocessing, model architecture, and model explainability approaches are detailed elsewhere (data not provided). Model architecture and training are detailed in the same, with final hyperparameters for both models after tuning of kernel size 17, batch size 32, learning rate 0.001. Model performance was evaluated using area under the receiver operating (AUROC) and precision recall (AUPRC) curves.RESULTS

[0386] The training and internal test cohorts comprised of 43,074 patients (median age at ECG 1.8 [IQR, 0.6-4.7] months) and 18,408 patients (median age at ECG 1.8 [IQR, 0.6-4.4] months), respectively, with CCHD prevalence of 21.3% and 21.1%, respectively. The external test cohort comprised of 162 patients, 25.9% with CCHD.

[0387] FIGS. 34A-C illustrate infant Electrocardiogram-Based Deep Learning (AI-ECG) to predict critical congenital heart disease. FIG. 34A illustrates performance of the infant artificial intelligence-enhanced electrocardiogram model evaluated during internal (blue) and external (orange) testing with receiver operating (AUROC; left) and precision-recall (AUPRC; right) curves. AUROC and AUPRC metric values for each model and outcome are inset. Dotted line represents chance (color-coded for differing prevalence in the precision-recall curve). 95% confidence intervals are shown using bootstrapping. FIG. 34B illustrates lesion-specific AUROC and AUPRC performance during internal (blue) and external (orange) testing. Color-coded prevalence of each lesion inset below. Bar graph represents mean ± standard deviation. FIG. 34C illustrates internal testing of lesion-specific AUROC (black) and AUPRC (red) performance. Prevalence of each lesion inset below. Bar graph represents mean ± standard deviation.Abbreviations: Positive predictive value (PPV); critical congenital heart disease (CCHD); complete atrioventricular canal defect (CAVC); coarctation of the aorta (CoA); double outlet right ventricle (DORV); D-loop transposition of the great arteries (D-TGA); pulmonary atresia (PA); total anomalous pulmonary venous connections (TAPVC); tetralog}' of Fallot (ToF); double inlet left ventricle (DILV); double inlet right ventricle (DIRV); double outlet left ventriclePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 (DOLV); hypoplastic left heart syndrome (HLHS); interrupted aortic arch (IAA); tricuspid atresia (TA); pulmonary stenosis (PS); aortic stenosis (AS).

[0388] Internal model performance to predict CCHD was similar when using 9- or 12- lead ECG inputs (p=0.9). For the 9-lead ECG model, performance is similar during internal testing (AUROC 0.94, AUPRC 0.83) and external validation (AUROC 0.92, AUPRC 0.82) (FIG. 34A). In newborns <7 days old (4.4% of cohort), performance remained high (AUROC 0.89, AUPRC 0.92). Performance to predict specific CCHD lesions is shown in FIG. 34B and FIG. 34C

[0389] FIG. 35 illustrates results for neonatal model explainability. Visualization of neonatal (< 7 days old) lesion-specific median waveforms generated in each lead using ECGs from the 25 highest (red) and 25 lowest (green) ALECG predictions. Saliency mapping demarcates regions of the ECG waveform having greatest (dark blue) and least (light blue) influence on predicting each lesion. Saliency was averaged over the 25 highest predicted ECGs for each lesion. Abbreviations: Complete atrioventricular canal defect (CAVC); coarctation of the aorta (CoA); pulmonary atresia (PA); tetralogy of Fallot (ToF); hypoplastic left heart syndrome (HLHS).

[0390] Model behavior analysis in neonates <7 days old (FIG. 35) demonstrated classical lesion-specific high-risk findings: rightward axis (tetralogy of Fallot [ToF]), superior axis (complete atrioventricular canal defect), right atrial enlargement with left ventricular hypertrophy (pulmonary atresia), right ventricular hypertrophy (CoA and HLHS), and decreased anterior right ventricular forces (tricuspid atresia). Across all lesions, QRS complexes were the most salient. DISCUSSION

[0391] Delayed diagnosis of CCHD remains a life-threatening issue in high- and low- resource settings, underscoring the need for inexpensive yet accurate CCHD diagnostic tools. Herein, the potential for AI-ECGto detect CCHD was demonstrated, notably with similar performance across institutions with vastly different demographics and good performance in lesions with low sensitivity during pulse oximetry CCHD screening (e.g., CoA, Ebstein anomaly, interrupted aortic arch, and ToF).

[0392] Practically, this tool may: 1) complement pulse oximetry as an enhanced two-part CCHD screening tool (to reduce late diagnoses); 2) inexpensively identify specific CCHD lesions (for low-resource settings).

[0393] Example 6: Artificial Intelligence-Based Electrocardiogram Predicts Sudden Cardiac Death in Repaired Tetralogy of FallotPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0394] The purpose of this example is to provide exemplary data and uses cases of a trained model described herein.

[0395] Despite progress in the management of patients with repaired tetralogy of Fallot (rTOF), ventricular arrhythmias and sudden cardiac death (SCD) remain challenging. Although numerous risk scores have attempted to identify patients who may benefit from an implantable cardioverter-defibrillator (ICD),2 the prediction of patients at risk for these rare yet lethal events remains unsatisfactory. Many risk stratification tools rely on resource-consuming modalities (e.g., cardiac magnetic resonance [CMR], invasive electrophysiology testing), prohibiting serial risk assessment at each clinic visit. The inventors have demonstrated the promise of inexpensive and ubiquitous artificial intelligence-based electrocardiogram (AI-ECG) to risk-stratify patients with rTOF. In this 2-center study, the potential for AI-ECG was explored to inform SCD risk stratification and associated ICD timing.

[0396] METHODS

[0397] A convolutional neural network was previously trained on ECGs obtained at Boston Children’s Hospital (BCH) and tested on nonoverlapping BCH and Toronto General Hospital (TGH) INDICATOR (International Multicenter Tetralogy of Fallot Registry) cohorts to predict 5-year all -cause mortality. Herein, the same AI-ECG algorithm were hypothesized to predict the likelihood of 5-year SCD, defined as a composite outcome of SCD or aborted SCD. Given the low outcome rate, the BCH and Toronto General Hospital cohorts were pooled into a single test set. Given the objective to inform primary prevention ICD timing, all ECGs after ICD placement were excluded. Area under the receiver operating curve (AUROC) was evaluated on AI-ECG probabilities to predict 5-year SCD. Model performance was benchmarked to QRS interval duration and CMR-derived biventricular global function index (BVGFI).4 AUROCs were compared via the DeLong test. To properly benchmark to imaging biomarkers, only ECG- CMR pairs of <1 year apart were included. To account for the variable number of ECGs per patient, AI-ECG performance was assessed by randomly selecting a single ECG recording per patient. Performance metrics were calculated through 1,000 iterations of this random sampling procedure. Patients were stratified based on previously established cutoffs — AI-ECG probabilities of >0.05 and BVGFI of <37%. Kaplan-Meier methodology was used to estimate outcome event rates on a single random ECG per patient. Patients who did not experience SCD were censored at the time of last known follow-up. Cox proportional hazards regression was used to calculate the C-statistic and evaluate risk group association with time from ECG until primary outcome occurrence. Institutional Review Board approval was obtained.

[0398] RESULTSPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0399] The pooled test cohort included 1,366 patients (17,159 ECGs at a median age of 23.5 [Q1-Q3: 11.4-37.6] years; 54% male; 20 [1.5%] patients experiencing the composite SCD outcome). The paired ECG-CMR test cohort included 1,245 patients (7,827 ECG-CMR pairs at a median age of 22.2 [Q1-Q3: 14.0-35.0] years; 56% male; 20 [1.6%] patients experiencing the composite SCD outcome). Median follow-up times after a single random ECG per patient in the pooled and paired cohorts were 7.5 and 8.0 years, respectively. AI-ECG probabilities (AUROC: 0.82; 95% CI: 0.75-0.89) outperformed QRS interval duration (AUROC: 0.71; 95% CI: 0.61- 0.81; P < 0.0001) and performed similarly to BVGFI (AUROC: 0.88; 95% CI: 0.82-0.93; P = 0.21) to predict the composite SCD outcome (FIG. 36A).

[0400] Stratifying by AI-ECG risk group led to good discrimination (C-statistics: 0.74; 95% CI: 0.64-0.85); compared with AI-ECG-predicted low-risk patients, AI-ECG-predicted high-risk patients were 9.5 times more likely to experience the primary outcome with significant differences (log-rank P < 0.0001) in freedom from SCD (FIG. 36B, left). Patients with 2 risk factors (paired AI-ECG probabilities of >0.05 and BVGFI of <37%) were at dramatically increased risk of SCD (15-year event rate of 29.7%) compared to patients with 0 or 1 risk factor (15-year event rates of 1.3%-2.8%) (FIG. 36B, right).

[0401] FIGS. 36A-B illustrate Performance in the (left) overall (n = 1,366; n = 20 composite SCD outcomes) and (right) paired ECG-CMR (n = 1,245; n = 20 composite SCD outcomes) test cohorts. (A) AUROC of AI-ECG to predict the 5-year risk of SCD. Model benchmarked to (left) QRS interval duration and (right) BVGFI (right). The dotted line represents chance, and the 95% Cis are shown in brackets. (B) (Left) Freedom from SCD Kaplan-Meier analysis when stratifying patients using AI-ECG only. Low-risk: AI-ECG prediction of <0.05 (green); high-risk: AI-ECG prediction of >0.05 (red). (Right) Freedom from SCD Kaplan-Meier analysis when stratifying patients based on the number of the following risk factors: Af-ECG prediction of >0.05 and BVGFI of <37%. HR (relative to the low-risk group), log-rank testing, number at risk, and event rates are inset below. AIECG = artificial intelligencebased electrocardiogram; AUROC = area under the receiver operating curve; BVGFI = biventricular global function index; CMR = cardiac magnetic resonance; ECG = electrocardiogram; SCD = sudden cardiac death.

[0402] DISCUSSION

[0403] Accurate SCD risk prediction to inform ICD timing has been a central question in rTOF management for decades. The historical use of QRS interval duration of >180ms and ECG accessibility motivated efforts to leverage the inventors’ previously developed AI-ECG all-cause mortality model to predict SCD. The AUROC achieved herein (0.82) outperforms QRS duration and is similar to BVGFI and a prior risk calculator (0.79) that notably requires CMR, exercisePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 stress test, and laboratory biomarker inputs. This AI-ECG algorithm can complement established frameworks for the consideration of primary prevention ICDs in rTOF. Demographics / clinical history, conventional ECG metrics, and ventricular health markers have been suggested to be aggregated to deem a patient as low risk (surveillance only), intermediate risk (invasive electrophysiological testing), or high risk (ICD placement). FIG. 36B illustrates a potential modified stepwise framework: AI-ECG low-risk patients may be followed expectantly with a low frequency of costly tests such as CMR. AI-ECG high-risk patients may require CMR; patients then found to have BVGFI of <37% may need invasive electrophysiologic testing and / or a primary prevention ICD. Future work includes expanded multicenter validation and prospective trials to inform clinical implementation.

[0404] Example 7: Electrocardiogram-Based Deep Learning to Predict Mortality in Repaired Tetralogy of Fallot

[0405] The purpose of this example is to provide exemplary data and uses cases of a trained model described herein.

[0406] Although advancements in complex congenital heart disease care have led to decreased early childhood mortality in patients with repaired tetralogy of Fallot (rTOF), these patients often experience residual hemodynamic and electrophysiologic abnormalities linked to increased rates of morbidity and premature mortality in adulthood. This trend underscores the need for robust risk stratification tools in this population. To date, risk factors have included conventional electrophysiologic (e.g., QRS duration and QRS fragmentation) and cardiac magnetic resonance (CMR)-based imaging biomarkers (including biventricular global function index [BVGFI]).

[0407] Other work has shown the promise of deep learning-based artificial intelligence- enhanced electrocardiogram (AI-ECG) algorithms for mortality prediction in adults with acquired cardiovascular diseases. Congenital heart lesions are associated with highly variable electrocardiogram (ECG) patterns (e.g., rTOF patients frequently have complete right bundle branch blocks), highlighting the need to develop congenital heart disease-specific AI-ECG algorithms. To date, there remains a paucity of AI-ECG applications to pediatric and adult congenital cardiology and no externally validated AI-ECG algorithm to predict mortality in rTOF.

[0408] In this work, this gap was addressed by training an AI-ECG model predictive of 5- year mortality on a diverse congenital and pediatric heart disease cohort at Boston Children’s Hospital (BCH). The model’s ability was then tested to predict 5-year mortality in patients with rTOF at BCH (internal testing) and Toronto General Hospital (TGH) (external validation).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 Finally, the additive value of the AI-ECG model was evaluated in predicting mortality to models based on traditional ECG and emerging imaging biomarkers.

[0409] METHODS

[0410] AI-ECG TRAINING COHORT

[0411] Training data was derived from all patients undergoing >1 ECG in the Cardiology Clinic at BCH between 1990 (earliest available electronically stored ECG) through June 2018. The end date was selected to minimize right censoring in predicting 5-year mortality as training data were retrieved from an internal database in November 2023. BCH patients enrolled in the INDICATOR (International Multicenter TOF Registry) study were removed from the training cohort and reserved for internal testing (FIG. 37).

[0412] INTERNAL TESTING AND EXTERNAL VALIDATION STUDY POPULATIONS.

[0413] BCH patients enrolled in the INDICATOR study comprised the internal testing cohort. Inclusion criteria for this group included: 1) enrollment in the INDICATOR cohort; 2) >1 ECG after repair between 2002 and December 2016 (end date selected to minimize right censoring given that INDICATOR was last updated December 31, 2021); and 3) documented follow-up or death after ECG.

[0414] The external validation cohort comprised rTOF patients from TGH enrolled in the INDICATOR cohort with >1 electronically stored ECG after repair with documented follow-up or death after ECG (FIG. 37). Similar to the internal test cohort, ECGs from November 2001 through December 2016 were used to assess 5-year mortality.

[0415] Data from BCH include pediatric and adult rTOF patients, whereas data from TGH include adult rTOF patients (i.e., >18 years of age) only.

[0416] DATA RETRIEVAL, QUALITY CONTROL, AND DATA PREPROCESSING.

[0417] Data retrieval, quality control, and data preprocessing methods have been previously described (data not provided). Briefly, retrieved waveform data corresponded to 1- dimensional vector of data sampled at 250 Hz for 10 seconds duration (2,500 samples) for a given lead. An ECG was discarded if any lead was not 2,500 samples long, or if any lead recording had no lead information (i.e., flat line). The ECG was then trimmed to 2,048 samples (w8 seconds) to facilitate conveniently working with convolution neural networks.

[0418] OUTCOMES.

[0419] The primary outcome was <5-year all-cause mortality after an ECG. As previously described, outcomes were ascertained by periodic searches of institutional electronic health records and searches of the National Death Index. Secondary outcomes included <1-, 2-, 3-, and 4-year mortality after an ECG. In Kaplan-Meier survival analysis for time-to-mortality, patients who did not experience the primary outcome were censored at the time of last follow-up.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0420] PRIMARY MODEL SELECTION, ARCHITECTURE, AND TRAINING.

[0421] We used 12 leads x 2,048 ECG samples as inputs to a convolutional neural network with a previously described architecture. The output of the last block is fed into a fully connected layer with a sigmoid activation function, given that the outcomes (<!-, 2-, 3-, 4-, and 5-year mortality after an ECG) are not mutually exclusive. To tune the AI-ECG model, the training set was further partitioned 95% for training and 5% for validation. The final hyperparameters (kernel size 9, batch size 64, learning rate 0.001) with lowest validation loss were obtained via a grid search on the following search space: kernel size (3, 9, 17), batch size (8, 32, 64), and initial learning rate (0.01, 0.001). The average cross-entropy was minimized using the Adam optimizer.

[0422] SECONDARY MODELS.

[0423] For comparison to the primary' model herein, secondary' rTOF-specifrc models were also generated on the internal test set using a: 1) 5-fold nested cross validation approach (by splitting the internal test set at the patient level into 5 folds, performing hyperparameter tuning on the training set for each fold, and evaluating performance on the holdout set; data not provided for further details); and 2) transfer learning approach (by starting with model weights from the primary model, and then training on the internal test set).

[0424] PERFORMANCE EVALUATION AND STATISTICAL ANALYSES.

[0425] Given the imbalanced dataset (i.e., low prevalence of mortality), both areas under the receiver operating curve (AUROC) and under the precision-recall (i.e., positive predictive value [PPV]- sensitivity) curve (AUPRC) were computed. To account for the variable number of ECGs per patient, AIECG performance was assessed by randomly selecting a single ECG recording per patient. Performance metric median and 95% Cis were calculated through 1,000- iterations of this random sampling procedure. To benchmark the primary model, prediction of 5- year mortality based on AI-ECG was compared to published rTOF risk stratification markers — QRS duration and BVGFI. To properly compare ECG and imaging risk stratification markers, ECGs and CMRs meeting the following criteria were paired: 1) ECG and CMR <1 year apart; 2) QRS duration and BVGFI available; 3) ECG-CMR pair occurrence before 2017; and 4) documented follow-up after ECG or event occurrence. Only the closest ECG-CMR pair was included, leaving a single ECG-CMR pair per patient. For this analysis, given the limited sample size, data from internal and external test cohorts were pooled together.

[0426] Other performance metrics evaluated included the model’s PPV, negative predictive value (NPV), sensitivity, and specificity. These metrics were calculated at the classification threshold achieving the maximal sum of sensitivity' and specificity (i.e., Youden index) in the training set.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0427] TRENDING AI-ECG PREDICTIONS OVER TIME.

[0428] As an exploratory analysis to assess the potential of AI-ECG to serve as a surveillance tool, internal and external AI-ECG 5-year mortality probabilities were assessed as a function of age at ECG in 1-year bins. The age at ECG was rounded to the nearest integer (i.e., 1- year bin), providing a range of AI-ECG probabilities within each bin. Within these 1-year bins, the median and interquartile range of AI-ECG probabilities were then reported.

[0429] SURVIVAL ANALYSIS.

[0430] Cox proportional hazards regression was used to evaluate factors associated with time from ECG until occurrence of all-cause mortality. A single ECG per patient was used in this analysis. Factors evaluated included AI-ECG output predictions of 5-year mortality, age at ECG, QRS duration, and BVGFI. Patients who did not experience death were censored at the time of last known follow-up. Kaplan-Meier methodology was used to estimate primary outcome event rates.

[0431] MODEL EXPLAINABILITY.

[0432] To explain model behavior, median waveform analysis and saliency mapping were performed as previously described (data not provided).

[0433] DATA A VAILABILITY AND SOFTWARE.

[0434] Requests for BCH data and related materials will be internally reviewed to clarify if the request is subject to intellectual property or confidentiality constraints. Shareable data and materials will be released under a material transfer agreement for noncommercial research purposes. Use of BCH and TGH data was approved by their respective institutional review boards.

[0435] Programming codes used to perform the analyses are described previously (data not provided).

[0436] RESULTS

[0437] TRAINING COHORT CHARACTERISTICS.

[0438] The demographic, clinical, and outcome characteristics of the training cohort are shown in TABLE 16. The cohort comprised 216,503 ECGs from 78,578 patients (median age at ECG, 10.6 [Q1-Q3: 3.4-17.0] years; 52% male) (FIG. 37). A wide range of pediatric and congenital heart diseases were included in the training cohort (TABLE 16). Of 78,578 patients, 1,618 (2.1%) died at a median age of 25.3 (Q1-Q3: 15.4-39.7) years. One- and 5-year mortality rates after ECG were 0.8% and 2.5%, respectively (TABLE 16). The training dataset included 1,736 (2.2%) patients with tetralogy of Fallot (TOF), none of whom were enrolled in the INDICATOR cohort (FIG. 37). A comparison between training TOF and internal test cohort baseline characteristics was performed (data not provided). The training TOF cohort was youngerPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 (median age at ECG: 12.2 years) with higher rates of mortality (8.9%) at a younger age (median age: 27.6 [Q1-Q3: 10.2-42.7] years).

[0439] TABLE 16: Baseline Characteristics of Training CohortDemographicsPatients 78,578SexFemale 37,607 (48)Male 40,960 (52)Unavailable 11 (<0.1)Mortality 1,618 (2.1)Median age at death, y 25.3 (15.4 - 39.7)DiagnosisVSD 8,219 (10)Cardiomyopathy 3,535 (4.5)ASD 2,984 (3.8)CoA 2,957 (3.8)D-loop TGA 1,782 (2.3)TOF 1,736 (2.2)Pulmonary atresia 1,082 (1.4)DORY 1,031 (1.3)HLHS 952 (1.2)L-loop TGA 767 (1.0)Dextrocardia 623 (0.8)Tricuspid atresia 513 (0.7)Ebstein 450 (0.6)TAPVR 479 (0.6)CAVC 443 (0.6)Truncus arteriosus 297 (0.4)ECGs 216,503Median age ECG 10.6 (3.4 - 17.0)Mortality <1 y of ECG 1,628 (0.8)Mortality <2 y of ECG 2,817 (1.3)Mortality <3 y of ECG 3,753 (1.7)Mortality <4 y of ECG 4,654 (2. 1)Mortality <5 y of ECG 5,488 (2.5)

[0440] Values are n (%) or median (Q1-Q3).

[0441] ASD = atrial septal defect; CAVC = complete atrioventricular canal defect; CoA= coarctation of the aorta; DORV = double outlet right ventricle; ECG = electrocardiogram;HLHS = hypoplastic left heart syndrome; TAPVR = total anomalous pulmonary' venous return;TGA = transposition of the great arteries; TOF = tetralogy of Fallot; VSD = ventricular septal defect.

[0442] CHARACTERISTICS OF THE INTERNAL TESTING AND EXTERNALVALIDATION COHORTS.

[0443] The internal testing and external validation cohorts comprised of 13,077 ECGs (n = 1,054) and 5,014 ECGs (n = 335), respectively (FIG. 37, TABLE 17), with notable differences in baseline characteristics. Compared with the BCH cohort, patients in the TGH cohort werePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 older at the time of TOF repair (median age: 5.2 [Q1-Q3: 3.8- 8.6] years vs 0.7 [Q1-Q3: 0.2-3. 1] years; P < 0.001) and at the time of baseline ECG (median age: 38.3 [Q1-Q3: 29.1-48.7] years vs 17.8 [Q1-Q3: 7.9-30.5] years; P < 0.001), and had a lower incidence of recorded genetic anomalies (10% vs 18%, P < 0.001). ECG-CMR pairs (n = 918 at BCH and n = 325 at TGH) were notable for older age, longer QRS duration, and increased body mass index in the external cohort, whereas median BVGFI was similar across cohorts (47.4% at BCH vs 47.1% at TGH; P = 0.70). Although prevalence of mortality was similar across internal testing and external validation cohorts (6.1% and 8.4%, P = 0. 14), there was a higher prevalence of mortality within 1- through 5-years after ECG for the external validation cohort (P < 0.001). Age at death was higher in the external validation cohort (52.4 [Q1-Q3: 43.2-62.9] years vs 39.9 [Q1-Q3: 19.8- 56.4] years; P = 0.003). There was a similar proportion of cardiac (w40%), noncardiac (w30%), and unknown (w30%) modes of death across cohorts (P = 0.8). Within 27 cardiac mortalities for the internal testing cohort, 7 (26%) were from sudden cardiac death, and 7 (26%) were from heart failure. Within 13 cardiac mortalities for the external validation cohort, 2 (15%) were from sudden cardiac death, and 7 (46%) were from heart failure.

[0444] TABLE 17: Baseline Characteristics of Training CohortPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0445] Values are n (%) or median (Ql-Q3).<

[0446] ASD = atrial septal defect; CAVC = complete atrioventricular canal defect; CoA = coarctation of the aorta; DORV = double outlet right ventricle; ECG = electrocardiogram; HLHS = hypoplastic left heart syndrome; TAPVR = total anomalous pulmonary' venous return; TGA = transposition of the great arteries; TOF = tetralogy of Fallot; VSD = ventricular septal defect.

[0447] AI-ECG MODEL PERFORMANCE.

[0448] FIG. 38A-B illustrate (A) Artificial intelligence-enhanced electrocardiogram (AI- ECG) model performance to predict 5-year mortality during internal testing (blue) and external validation (orange). Area under the receiver operating curve (AUROC) and area under the precision-recall curve (AUPRC) metric values for each model are inset. Dotted line represents chance. 95% Cis are shown using bootstrapping. (B) Kaplan-Meier curve survival analysis of internal (left) and external (right) cohorts when stratifying patients as low- (green) or high-risk (red) based on AI-ECG predictions. Statistics based on log-rank testing. Number at risk within each group is shown in below.

[0449] Model performance to predict 5-year mortality was similar during internal testing (AUROC: 0.83, AUPRC: 0.18) and external validation (AUROC: 0.81, AUPRC: 0.21) (FIG. 38A). Similar model performance was achieved in a sensitivity analysis when excluding noncardiac mortality as an outcome (data not provided). NPVs of 98% to 99% and PPVs of 11% to 14% were achieved across cohorts (data not provided).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0450] Using the Youden index threshold of 0.05 (AI-ECG prediction range: 0-1) to distinguish between patients at low- (prediction <0.05) vs high-risk (prediction >0.05) of 5-year mortality, the internal testing and external validation cohorts had similar percentages of low-risk patients (75% to 77%). Survival analysis showed effective discrimination in both cohorts (both P < 0.0001) (FIG. 38B). In Cox models to predict time to death, the AI-ECG model 5-year mortality predictions (internal testing and external validation c-index of 0.81 [range: 0.76- 0.87] and 0.81 [range: 0.72-0.89], respectively) outperformed age at ECG (internal testing and external validation c-index of 0.73 [range: 0.65-0.81] and 0.73 [range: 0.63-0.82], respectively) (data not provided). Adding age at ECG to the AI-ECG predictions provided no benefit to predict shorter time to death (data not provided).

[0451] rTOF-SPECIFIC MODELS.

[0452] As secondary analyses, model performance was explored when training secondary rTOF-specific models by: 1) performing 5-fold nested cross-validation on the internal test cohort; and 2) retraining the primary model on the internal test cohort. During 5-fold nested cross- validation, internal test performance was similar (AUROC: 0.78 [95% CI: 0.74-0.83] and AUPRC: 0.19 [95% CI: 0.13-0.26]) to the primary model herein. The retrained model performed worse (AUROC: 0.74 [95% CI: 0.65-0.82], AUPRC: 0.18 [95% CI: 0.07-0.33]) than the primary model on the external validation cohort.

[0453] BENCHMARKING AI-ECG MODEL PERFORMANCE.

[0454] AI-ECG performance in predicting the primary outcome was benchmarked to established rTOF risk stratification markers — QRS duration and BVGFI. AI-ECG (AUROC: 0.86, AUPRC: 0.23) outperformed QRS duration (AUROC: 0.69, AUPRC: 0.08) and performed similarly to BVGFI (AUROC: 0.82, AUPRC: 0.17) (FIG. 39A).

[0455] Survival after ECG-CMR pair was assessed by stratifying patients into low- prediction <0.05) or high-risk (prediction >0.05) AI-ECG groups (FIG. 39B), as well as low- (BVGFI >37) or high-risk (BVGFI <37) BVGFI groups. For both AI-ECG- and BVGFI-based stratifications, similarly effective discrimination of survival rates was achieved in both internal testing and external validation cohorts (P < 0.0001).

[0456] Among Cox models (TABLE 18) formulated with individual predictors (BVGFI, QRS duration, or AI-ECG 5-year mortality predictions), AI-ECG 5-year mortality predictions achieved the highest c-index during internal testing and external validation (0.84 and 0.77, respectively). When combining conventional imaging biomarkers and ECG parameters (i.e., BVGFI [) QRS duration), QRS duration was no longer a significant predictor (P = 0.20-0.43) (TABLE 18). In contrast, AI-ECG predictions remained a significant outcome predictor whenPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 combined with BVGFI for internal testing and external validation cohorts with the highest c- index (0.86 [95% CI: 0.81-0.91] and 0.81 [95% CI: 0.73-0.93], respectively) (TABLE 18).

[0457] TABLE 18: Benchmarking Cox Model Performance During Internal and ExternalTesting

[0458] SUBGROUP ANALYSIS.

[0459] In a subgroup analysis (data not provided), model performance was similar across a wide range of categories such as demographics (e.g., age), procedure history , BVGFI, and arrhythmia history; however, model performance was lower for patients with genetic syndromes. Performance was similar between patients with (AUROC: 0.83, AUPRC: 0.14) and without (AUROC: 0.83, AUPRC: 0.18) complete right bundle branch blocks.

[0460] AI-ECG ASA SURVEILLANCE TOOL.

[0461] The relationship between age at ECG and AI-ECG probabilities of 5-year mortality changed exponentially with increasing age (FIG. 40). AI-ECG probabilities of 5-year mortality remained low during the first 3 decades of life, followed by an exponential increase beginning during the fourth decade onwards.

[0462] FIG. 40 illustrates Internal testing (left) and external validation (right) AI-ECG probabilities of 5-year mortality (y-axis) as a function of age at ECG (x-axis). Median (solid line) with interquartile range (shaded region) of AI-ECG probabilities for each 1-year age bin shown. Abbreviations as in FIGS. 37-38.

[0463] MODEL EXP LAIN ABILITY.

[0464] FIG. 41 illustrates Visualization of median waveforms generated in each lead using ECGs from the 50 highest (red) and 50 lowest (green) AI-ECG predictions of the internal (top) and external (bottom) cohorts. Saliency mapping highlights ECG regions with greatest (dark blue) and least (light blue) influence on mortality . Saliency was averaged over the 50 highest predicted ECGs. Abbreviations as in FIGS. 37-38.

[0465] Saliency mapping and median waveform analysis were performed to interpret model behavior (FIGS. 41A-B) In both internal testing (FIG. 41A) and external validation (FIG. 41B) cohorts, salient features included precordial QRS complexes (VI, V3, V6) and TPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 waves (leads II, aVR, VI, and V6). High-risk features included wide- and low-amplitude QRS complexes (all leads), QRS fragmentation (V2-V4), flattened T waves (lateral precordial leads), and pro- longed terminal repolarization (lead II and aVR).

[0466] DISCUSSION

[0467] In this work, an AI-ECG algorithm was developed and externally validated to predict mortality in rTOF (FIG. 42). The model achieved similar performance during internal testing and external validation on cohorts with notably different baseline characteristics, suggesting model generalizability . AI-ECG outperformed QRS duration and performed similarly to CMR-based BVGFI, providing a convenient and accessible risk- stratification tool that may complement imaging biomarkers. Model performance was consistent across a range of subgroups (data not provided). Finally, saliency mapping and median waveform analysis provided a framework to develop insights into clinically relevant ECG characteristics predictive of mortality, including QRS fragmentation. Altogether, these findings demonstrate the promise of Al- ECG to inexpensively screen for mortality in rTOF, which may help rationalize use of expensive imaging studies and democratize access to specialized care in this growing group of patients.

[0468] TRADITIONAL rTOF RISK STRATIFICATION

[0469] Among the many parameters identified as being associated with adverse clinical outcomes in rTOF, CMR-based imaging biomarkers and ECG metrics have often been used for risk stratification. Among the various CMR parameters, right ventricular size and left and right ventricular systolic function have been shown to be associated with adverse clinical outcomes. Later studies reinforced the central role of CMR — and the value of biventricular dysfunction, BVGFI, and late gadolinium enhancement — to predict mortality. In this study, an AI-ECG algorithm was found to be as useful as CMR-based BVGFI to predict 5- year mortality in patients with rTOF, achieving a similar c-index to a recent CMR-based risk score during external validation.

[0470] ECG similarly plays a central role in the surveillance of patients with rTOF. Importantly, although QRS prolongation (QRS duration >180 ms)17 and fragmentation have been previously identified as mortality risk factors, they have been generally overshadowed by CMR metrics in contemporary guidelines. In this study, AL ECG was shown to outperform conventional ECG metrics (i.e., QRS duration) and performs similarly to CMR metrics (i.e., BVGFI), providing an easily accessible yet powerful prognostic tool for predicting mortality in patients with rTOF.

[0471] CLINICAL SIGNIFICANCE AND IMPLICATIONS.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0472] CMR has practical limitations (e.g., time-intensive, resource-consuming) hindering its widespread use. In contrast, ECG circumvents these limitations and can be obtained at every cardiology visit, facilitating easier and more frequent use. This AI-ECG algorithm may serve as a screening and surveillance tool, which may also improve access to care. As a screening tool, this algorithm may have eco- nomic value by reducing the frequency of imaging tests (e.g., echocardiography, CMR). At the given threshold of 0.05 (range 0-1), approximately 75% of patients were deemed low risk with a 5-year NPV of 98% to 99% across cohorts. Current guidelines recommend CMR surveillance every 24 to 36 months in rTOF adults with stage A-B physiologic classification. Theoretically, for patients in this low-risk category, without any clinical change, it may be reasonable to decrease the frequency of CMRs to every 5 years. On the other hand, AI-ECG predictions could help identify high-risk patients within the stage A-B classification at an earlier age who will require closer monitoring and earlier CMR studies to consider therapeutic interventions such as pulmonary valve replacement or implantable cardioverter- defibrillator. Other work has shown that QRS fragmentation (and not guideline- recommended risk factors) was the only independent predictor of appropriate implantable cardioverter-defibrillator therapy for patients with TOF. Similarly, QRS duration has been used for conservative / proactive consensus criteria for pulmonary valve replacement. Further work is necessary to investigate if Al- ECG output predictions may serve a similar purpose to guide decisions about therapeutic interventions.

[0473] As a surveillance tool, a congenital cardiologist could plausibly monitor AI-ECG predictions at each cardiology visit. In the exploratory population-level study (FIG. 40), there were similar increases in Al- ECG 5-year mortality probabilities with aging across institutions. Using AI-ECG to track responsiveness to interventions has shown promise in patients with symptomatic obstructive hypertrophy cardiomyopathy receiving mavacamten. Future work is necessary to assess whether a similar application is possible for tracking response to pulmonary valve replacement or other interventions.

[0474] Finally, if broadly validated, this algorithm may help improve access to care, especially for centers with limited access to CMR. Although the majority of congenital heart disease patients are now adults, approximately half are without regional congenital heart disease services. This technology may contribute to democratization of specialty expertise, and be informative for health care providers with limited expertise in congenital cardiology by reducing the need for specialized CMR knowledge to risk stratify rTOF.

[0475] CHOICE OF TRAINING COHORT.

[0476] Congenital heart disease and cardiomyopathy have nontnvial overlap, both of which contribute to pediatric heart failure. In this study, predicting mortality from a diversePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 congenital heart disease cohort was hypothesized to be informative for predicting rTOF specifically. This benefit, along with the significantly larger sample size, prompted us to train a generalized mortality model rather than a rTOF-specific mortality model. Generating rTOF- specific mortality models was found to provide no added benefit, which may be attributable to a smaller sample size.

[0477] Our primary model provides several other advantages compared to the secondary rTOF-specific models, including the potential to be more generalizable across a range of congenital heart lesions, warranting future investigation that is outside the scope of this study.

[0478] MODEL EXPLAINABILITY AND ECG INSIGHTS.

[0479] Model explainability provides transparency and interpretability for clinicians, which may aid in identifying ECG myopathy signatures. Previous rTOF-specific ECG signatures may be pertinent to the explainability findings herein. For example, QRS complexes were salient throughout, and it has been previously that identified early VI QRS signals correspond to right ventricular septal activation, whereas later portions represent right ventricular free wall, outflow tract, and basolateral activation. In the external cohort, terminal S waves are high-risk features noted in lead aVF, which may represent septal isthmus conduction abnormalities. Finally, leads II and aVR have high-risk features of prolonged terminal repolarization; failing cardiomyocytes have prolonged terminal repolarization related to ion channel remodeling, which may increase arrhythmia susceptibility. Arrhythmic risk may be further amplified by spatial dispersion of repolarization across the myocardium that may be reflected by non-uniform T wave changes. Future work is needed to rigorously correlate these hypothesis-generating findings to the underlying pathophysiology.

[0480] RELATIONSHIP BETWEEN AGE AND AI-ECG PREDICTIONS.

[0481] Previous work has shown that AI-ECG can directly predict patient age, and that the AI-ECG predicted biological age gap (i.e., AI-ECG predicted age minus actual age) is predictive of mortality. Herein, the model was similarly hypothesized to be learning age in addition to other pertinent rTOF hemodynamic / electrophysiologic abnormalities given: 1) there is a relationship noted between AI-ECG probabilities and age (FIG. 40); and 2) the addition of age to AI-ECG provides no benefit to model performance (data not provided). AI-ECG outperformed age to predict mortality (TABLE 18, complete data not provided), demonstrating that the model is learning more than age alone.

[0482] STUDY LIMITATIONS.

[0483] First, all-cause mortality was chosen as the primary outcome as opposed to cardiac death. This choice is consistent with several other AI-ECG works, as well as numerous investigations on outcomes in adult congenital heart disease. The distribution of causes of deathPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 (cardiac, noncardiac, and unknown) in the INDICATOR cohort is similar to previous reports in patients with congenital heart disease. In addition, performance remained high when excluding noncardiac death during internal testing and external validation (data not provided). Second, only one example of thresholding was used to evaluate model performance. Future consideration is needed to weigh the impact of resultant false positives / negatives. Multicenter external validation is required to further refine thresholds for clinical implementation. Similarly, multicenter collaboration may further increase training / testing sample sizes and may lead to a more diverse model with better performance. Third, the lack of an automated and reliable tool to detect QRS fragmentation in rTOF prohibited benchmarking with this known risk factor. Similarly, lack of late gadolinium enhancement prevents benchmarking to a recent model. Other risk factors worthy of bench- marking and / or incorporating into a model include, but are not limited to, B-type natriuretic peptide and assessment of fibrosis by CMR. Finally, the limitations of saliency mapping are noted.

[0484] FUTURE DIRECTIONS.

[0485] Future work may include: 1) expanded multicenter collaboration; 2) investigating the utility of serial ECG inputs (rather than single ECG inputs) and / or multimodal inputs; and 3) prospective trials to inform clinical implementation. Furthermore, fine-tuning of the model to other clinically relevant, disease-specific predictions (e.g., timing of pulmonary valve replacement; appropriate implantable cardioverter defibrillator shocks) may be of value.

[0486] CONCLUSIONS

[0487] This externally validated algorithm shows promise to inexpensively, accurately, and conveniently risk stratify patients with rTOF. ALECG may help rationalize use of costly imaging testing, provide insight into ECG signatures of mortality, and potentially reduce disparities by improving access to care. Future multicenter collaboration and prospective trials are warranted.

[0488] Example 8: Electrocardiogram-based deep learning to predict left ventricular systolic dysfunction in pediatric and adult congenital heart disease

[0489] The purpose of this example is to provide exemplary data and uses cases of a trained model described herein.

[0490] INTRODUCTION

[0491] With recent medical and surgical advancements, most children bom with congenital heart diseases (CHDs) now survive into adulthood. In the USA alone, it is estimated that approximately 2.4 million individuals are living with CHD, including approximately 1 million children and 1.4 million adults. These conditions require lifelong follow-up involvingPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 frequent clinic visits and cardiac testing with significantly higher associated costs throughout the lifespan. The necessity to facilitate the management of this growing complex population with a wide range of structural anomalies and distinct long-term consequences has motivated efforts to risk stratify CHD patients across the lifespan. Doing so may reduce a significant portion of unnecessary and expensive testing, while simultaneously identifying individuals that might benefit from closer monitoring or intervention thereby improving clinical outcomes.

[0492] However, the development of robust risk prediction models — particularly in CHD — faces multiple challenges including a paucity of big data, limited data sources, reliance on expensive imaging modalities (with an associated need for pediatric cardiologist expertise), and sub- optimal model performance. This premise motivated multiple recent calls for artificial intelligence applications to improve risk stratification in CHD.

[0493] Major adverse outcomes in this unique population often relate to ventricular dysfunction or arrhythmias, and extraction of conventional electrocardiogram (ECG) features such as QRS duration in CHD [e.g. tetralogy of Fallot (ToF)] has shown value in risk stratification. In addition, deep learning-based artificial intelligence-enhanced ECG (AI-ECG) algorithms show promise for diagnostic and prognostic applications in adults, making it similarly conceivable that AI-ECG may aid risk stratification also in the CHD population. However, there remains a paucity of available AI-ECG applications to congenital cardiology given (i) the absence of extensive data sets, impeding similar research; (ii) AI-ECG algorithms derived from adults with structurally normal hearts would be expected to have poor generalizability to CHD cohorts. To this end, there remain no AI-ECG algorithms to predict mortality in CHD.

[0494] In this work, an exhaustive electronic database at a large congenital heart center was leveraged to address this gap. To do so, a convolutional neural network was trained to predict 5-year mortality using >100 000 ECGs on nearly 40 000 patients and tested on an equally sized cohort as well as on a contemporary cohort. Subgroup analysis defined model performance within a range of specific CHDs. Survival analysis investigated longer-term survival for patients. Finally, saliency mapping and median waveform analysis provided lesion-specific model explainability.

[0495] METHODS

[0496] Our study adheres to the TRIPOD + Al guidelines.

[0497] Study population and patient assignment

[0498] Patient data were utilized from the Boston Children’s Hospital be- tween 1990 (the earliest available digitized ECG in the clinic) until June 2018 (to minimize right censoring in predicting 5-year mortality). Inclusion criteria consisted of any patient presenting to the cardiology7clinic at the Boston Children’s Hospital with at least one ECG per- formed. TemporalPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 validation for predicting 1-year mortality was performed on any patient presenting to the cardiology clinic at the Boston Children’s Hospital with at least one ECG performed from July 2018 to July 2022 (to minimize right censoring in predicting 1-year mortality). All data were retrieved from an internal database in November 2023. Similar to prior work, a group stratified design was used to partition the main cohort at the patient level such that all ECGs for a given patient were restricted to either the training or internal testing cohort. Patients were randomly partitioned 50:50 into training and internal testing cohorts. Patients from the training data set were excluded from the temporal validation cohort.

[0499] Data retrieval

[0500] Raw ECG signals were exported from the MUSE ECG data management system (GE Healthcare, Chicago, IL, USA), which contains ID vectors of data (sampling rate of 250 Hz for 10-s duration) for each lead (I, II, and Vl-6). From these vectors, Einthoven’s law and the Goldberger equation were implemented to obtain leads III, aVF, aVL, and aVR. Other data retrieved from an internal database at the Boston Children’s Hospital include age, sex, and physician-reviewed ECGs measurements (e.g. QRS interval, QRS axis, T axis, P axis, PR interval, QT interval, QTc, and heart rate).

[0501] Patients with known CHD, cardiomyopathy, pre-excitation syndromes, and channelopathies were identified based on the institutional Fyler coding system, which has been mapped into the International Paediatric and Congenital Cardiac Code International Classification of Diseases-11 nomenclatures.

[0502] Quality control and data pre-processing

[0503] An ECG was discarded if any lead was not 2500 samples long or if any lead recording had missing lead information. After quality control, a high pass filter was utilized to account for recording errors (e.g. baseline wander and electncal interference) with cut-off frequency 0.8 Hz, rejection band 0.2 Hz, ripple in passband 0.5 dB, and attenuation in rejection band 40 dB. Finally, the ECG was then trimmed to 2048 samples (~8 s) to facilitate conveniently working with convolutional neural networks.

[0504] Definition of outcomes

[0505] The primary outcome was 5-year mortality after an ECG. All-cause mortal- ity was used as the primary outcome given (i) heart failure is the most common cause of death in adult CHD and (ii) it aligns with previous ALECG work and recent CHD work. To obtain allcause mortality, date of death was retrieved from an internal institutional database. Secondary' out- comes include 1-, 2-, 3-, and 4-year mortality after an ECG. For temporal validation, 1-year mortality was used to minimize right censoring.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0506] To perform a secondary Kaplan-Meier survival analysis for time to mortality, the date last known alive within the internal database (i.e. last recorded event in the institutional database for a given patient) was obtained.

[0507] Model selection, architecture, and training

[0508] This AI-ECG model was developed solely on the training set, which was further partitioned 95% for training and 5% for validation to perform hyperparameter tuning. 12 x 2048 ECG samples were used as inputs to a convolutional neural network that is similar to the residual network previously described that is adapted for unidimensional signals. A diagram and details of the architecture used in this study are shown previously (data not provided). The output of the last block is fed into a fully connected layer with a sigmoid activation function as the outcomes (1-, 2-, 3-, 4-, and 5-year mortality) within this single model are not mutually exclusive.

[0509] The final hyperparameters were obtained via a grid search on the training set among the following options: kernel size [3, 9, 17], batch size [8, 32, 64], and initial learning rate [0.01, 0.001, 0.0001, 0.00001], The average cross- entropy was minimized using the Adam optimizer. Maximum 150 epochs were used with early stopping based on validation loss. The model wdth the lowest validation loss during hyperparameter tuning was selected as the final model. The final hyperparameters were kernel size 17, batch size 32, and learning rate 0.001.

[0510] Performance evaluation and statistical analyses

[0511] Given the potential for loss to follow-up, model performance for all binary outcomes (e.g. 1- and 5-year mortality) was assessed only on ECGs with documented follow-up after the outcome timeframe or mortality events within the outcome timeframe. Consistent with prior works, multiple ECGs per patient were allowed in the training cohort. In contrast, model performance was evaluated on one ECG per patient. The ECG selected for testing was either the first available, the last available, or a randomly selected ECG per patient.

[0512] Given the imbalanced data set (i.e. low prevalence of mortality), the area under the receiver operating characteristic curve (AUROC) and area under the precision-recall [i.e. positive predictive value (PPV) sensitivity] curve (AUPRC) were computed. To benchmark the model, age at ECG (previously used to benchmark AI-ECG predictions of mortality), QRS duration (a conventional ECG predictor of mortality in ToF), and QTc duration (an independent risk factor for sudden cardiac death) were used. In addition, left ventricular ejection fraction (LVEF) — an established determinant of cardiovascular morbidity and mortality — from an echo within 2 days of the paired ECG was used as a benchmark when available. Other performance metrics evaluated included PPV, negative predictive value (NPV), sensitivity, and specificity. These metrics w ere calculated at the classification threshold maximizing the sensitivity and specificity (i.e. the Youden index) in the train- ing set. For all metrics, a higher value is indicativePCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 of better performance. Resampling with 1000 bootstraps was implemented to obtain performance metric median and 95% confidence intervals (Cis). Area under the receiver operating characteristic curves were compared when applicable via the DeLong test.

[0513] Subgroup analyses

[0514] Subgroup analyses were performed on the test set for each disease of inter- est. Given that conduction disturbances are also a marker of disease progression in CHD, model performance was also assessed when stratifying by QRS duration and presence of complete right bundle branch block. Area under the receiver operating characteristic curves and AUPRCs were calculated for each subgroup.

[0515] Survival analysis

[0516] Cox proportional hazards regression was used to evaluate factors associated with time from ECG until all-cause mortality. Patients who did not experience death were censored at the time of last known follow-up. Patients with unknow n follow-up time after ECG were excluded from this secondary analysis. The Kaplan-Meier methodology was used to estimate primary outcome event rates.

[0517] Model explainability

[0518] In an effort to interpret model behavior, analyses of median waveforms and saliency mapping were performed. Median w aveform analysis is a technique to visually represent single beats of the highest and lowest risk ECGs. Herein, a subset of test set ECGs (100 for the overall cohort, 25 for specific lesions) with the highest predicted probability for a given outcome were used to create high-risk median waveforms, and a corresponding set with the lowest predicted probabilities was used to create low-risk median waveforms. Median waveforms were generated in each lead using the NeuroKit Python toolbox by (i) detecting QRS complexes, (ii) interpolating all ECGs to the same heart rate, (iii) computing the median voltage across beats for each patent, and (iv) computing the median voltage across patients for each time bin in the cardiac cycle. Saliency mapping aims to identify which features of the ECG input contribute to model prediction by highlighting components of the ECG where a change in input (i.e. ECG voltage) leads to a relatively large change in prediction. Saliency maps were created using a Shapley Additive Explanations (SHAP) framework for the high-risk ECGs. The above median waveform steps were similarly implemented on SHAP values over time. Darker regions in saliency maps correspond to greater contribution to the prediction of 5-year mortality.

[0519] Software

[0520] Programming codes used to perform the analyses are available upon rea- sonable request. The convolutional neural network used the Keras framework with a TensorFlow (Google) backend using Python 3.9.36 Deep learning was executed on institutional graphicsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 processing units. All other pre- and post-processing codes were written in Python 3.936 and R 4.0,37 which was executed locally.

[0521] RESULTS

[0522] Internal patient population characteristics

[0523] The training cohort comprised of 112 804 ECGs from 39 784 patients [median age at ECG 7.7 [interquartile range (IQR), 1.6-14.8; range 0-85] years; 52% male]. The first digitized ECG included in this cohort was from 1990. As shown in TABLE 19, a wide range of CHD lesions were included, including 11% with ventricular septal defects, 4.5% with cardiomyopathy, 3.8% with coarctation of the aorta, 3.5% with ToF, 2.2% with D-loop transposition of the great arteries, 1.2% with hypoplastic left heart syndrome, 1.0% with L-loop trans- position of the great arteries, 0.8% with dextrocardia, and 0.7% with tricuspid atresia. The internal testing cohort included 112 575 ECGs from 39 784 distinct patients [median age at ECG 7.9 (IQR, 1.5-14.8; range 0-92) years; 52% male] and composed of a similar breakdown of congenital heart lesions. The contemporary cohort included 42 927 ECGs from 25 537 patients. Numerous differences in baseline characteristics were noted in the contemporary cohort including higher prevalence of all disease groups (except ventricular septal defects) and older age at ECG [median age at ECG 10.5 (IQR, 3.1-16.7) years],

[0524] TABLE 19 Internal cohort baseline characteristicsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025coarctation of the aorta; CAVC, complete atrioventricular canal defect; CRBBB, complete right bundle branch block; DORV, double outlet right ventricle; HLHS, hypoplastic left heart syndrome; LVEF, left ventricular ejection fraction; PA, pulmonary atresia; ToF, tetralogy of Fallot; TAPVR, total anomalous pulmonary venous return; TGA, transposition of the great arteries; VSD, ventricular septal defect; WPW, Wolff-Parkinson- White.aNote select lesions included, some of which overlap.bDenominator is equal to the total number of ECGs with either mortality events within the outcome timeframe or with documented follow-up after the outcome timeframe.

[0526] In the training cohort, there were 806 (2.0%) mortality events at median age 18.9 (IQR, 8.0-32.4) years. In the testing cohort, there were 870 (2.2%) mortality events at median age 19.4 (IQR, 7.5-32.2) years. Five-year mortality after ECGs in the training and test cohorts was 4.9% and 4.6%, respectively. Similar 1-year mortality events (1.0%) were noted across internal test and temporal validation cohorts, with a younger age of mortality in the temporal validation cohort [median age 13.1 (IQR, 4.7-22.6) years]. Electrocardiograms with mortality within 5 years had higher heart rates, longer QRS and QTc intervals, and lower paired echo ejection fraction (data not provided).

[0527] FIGS. 43A-C illustrate Kaplan-Meier survival analysis of training and testing cohorts. (A) Kaplan-Meier curve survival analysis of training (green) and testing (red) cohorts demonstrates the large, diverse cohorts across the congenital heart disease lifespan. Lesionspecific survival curves for (B) left ventricular and (C) right ventricular pathology are shown when pooling training and testing cohorts. Number at risk within each group inset below. CoA, coarctation of the aorta; DORV, double outlet right ventricle; HLHS, hypoplastic left heart syndrome; LV, left ventricular; PA, pulmonary atresia; RV, right ventricular; TAPVR, total anomalous pulmonary venous return; TGA, transposition of the great arteries; ToF, tetralogy of Fallot.

[0528] Survival in the training and testing cohorts was similar [hazard ratio 1.0 (95% CI 0.9-1.0), P = .3; log-rank P = .2] and spanned across the life- span (FIG. 43A). As shown FIGS.PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 43B-C, a large portion of patients survive to adulthood, with survival rates varying by lesion. The lowest long-term survival was lesions that typically involve single ventricle palliation (tricuspid atresia for RV pathology, hypoplastic left heart syn- drome for left ventricular pathology).

[0529] Model performance

[0530] FIGS. 44A-C illustrate ECG-based deep learning model performance. (A) Artificial intelligence-enhanced electrocardiogram model performance to predict 5-year mortality evaluated using a random (blue), last (orange), and first (green) electrocardiogram per patient.(B) Performance benchmarking of the random electrocardiogram (blue) to age at electrocardiogram (orange), QRS duration (green), QTc duration (red), and left ventricular ejection fraction (purple). (C) Comparison of internal testing (blue) and temporal validation (orange) performance to predict 1-year mortality. Area under the receiver operating characteristic curve and area under the precision-recall curve metric values for each model and outcome are inset. Dotted line represents chance. 95% confidence intervals are shown using bootstrapping.AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve; ECG, electrocardiogram; LVEF, left ventricular ejection fraction; PPV, positive predictive value.

[0531] During testing (FIG. 44A), the model achieved the following performance in 5- year mortality when using the random, last, and first ECGs: AUROCs of 0.79 (95% CI 0.77- 0.81), 0.82 (95% CI 0.80-0.83), and 0.75 (95% CI 0.72-0.77), respectively, and AUPRCs of 0.17 (95% CI 0.15-0.19), 0.25 (95% CI 0.23-0.28), and 0.11 (95% CI 0.09-0.12), respectively. The randomly selected ECG performance outperformed age at ECG [AUROC 0.58 (95% CI 0.55- 0.61); AUPRC 0.11 (95% CI 0.08-0.13); P < .001], QRS duration [AUROC 0.57 (95% CI 0.54- 0.60); AUPRC 0.07 (95% CI 0.06-0.08); P < .001], QTc duration [AUROC 0.48 (95% CI 0.45- 0.50); AUPRC 0.05 (95% CI 0.04-0.06); P < .001], and paired echo LVEF [AUROC 0.62 (95% CI 0.57-0.68) [AUPRC 0.10 (95% CI 0.07-0.15); P < .001] when predicting 5-year mortality (FIG. 44B), as well as 1-year mortality (data not provided). Artificial intelligence-enhanced ECG similarly outperformed these metrics when using the last available ECG (data not provided).

[0532] When using a random ECG, the sensitivity, specificity, and PPV were 0.66 (95% CI 0.62-0.70), 0.78 (95% CI 0.78-0.79), and 12.4% (95% CI 11.6%-13.1%), respectively, with 76.4% (95% CI 75.6%-77. 1%) predicted negative (data not provided). A NPV of ~98% was achieved independent of ECG used (data not provided ). Model calibration is not provided.

[0533] Contemporary model performance

[0534] During temporal validation, AUROC of 0.79 (95% CI 0.74-0.83) and AUPRC of 0.04 (95% CI 0.03-0.06) was achieved to predict 1-year mortality (FIG. 44C).

[0535] Subgroup analysisPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0536] FIG. 45 illustrates model performance in congenital heart disease subgroups. Forest plot showing artificial intelligence-enhanced electrocardiogram area under the area under the receiver operating characteristic curve (red) and area under the precision-recall curve (black) performance when stratifying by lesion when using a random (left), last (middle), and first (right) electrocardiogram. Area under the receiver operating characteristic curve and area under the precision-recall curve metric values for each model and outcome are inset. 95% confidence intervals are shown using bootstrapping. ASD, atrial septal defect; CoA, coarctation of the aorta; CAVC, complete atrioventricular canal defect; DORV, double outlet right ventricle; ECG, electrocardiogram; HLHS, hypoplastic left heart syndrome; PA, pulmonary atresia; ToF, tetralogy of Fallot; TAPVC, total anomalous pulmonary venous connection; TGA, transposition of the great arteries; VSD, ventricular septal defect.

[0537] In a subgroup analysis (FIG. 45), model performance appeared lesion and age dependent. When using a random ECG per patient, model performance in extra-cardiac pathophysiology (i.e. coarctation of the aorta and total anomalous pulmonary venous return) was as follows: AUROCs of 0.86 (95% CI 0.81-0.91) and 0.80 (95% CI 0.72-0.88), respectively, and AUPRCs of 0.30 (95% CI 0.19-0.41) and 0.38 (95% CI 0.20-0.56), respectively. For a wide range of congenital heart lesions with accumulating pathophysiologic myocardial burden over time (e.g. cardiomyopathy, pulmonary atresia, double outlet right ventricle, hypoplastic left heart syndrome, and tricuspid atresia), lower AUROC with higher AUPRC patterns were noted. These trends were consistent when using the last available ECG, and less so when using the first available ECG. Similarly, the dextrocardiac subgroup had a lower AUROC [0.71 (95% CI 0.61— 0.81)] with higher AUPRC [0.21 (95% CI 0.10-0.32)]. Model performance was lower in L-loop transposition of the great arteries (FIG. 45). Across all lesions with sufficient data available for comparison, AUROC and AUPRC trended higher for AI-ECG compared with LVEF (data not provided).

[0538] Given that conduction disturbances are also a marker of disease progression in CHD, model performance was also assessed when stratifying by QRS duration and presence of complete right bundle branch block. In patients with QRS duration < 120 ms, AUROC of 0.77 (95% CI 0.75-0.80) and AUPRC of 0.14 (95% CI 0.12-0.16) were achieved. For QRS duration > 120 ms, AUROC of 0.76 (95% CI 0.72-0.80) and AUPRC of 0.20 (95% CI 0.15-0.24) were obtained. In patients with complete right bundle branch block, AUROC of 0.78 (95% CI 0.73- 0.83) and AUPRC of 0.21 (95% CI 0.13-0.29) were achieved. In patients without complete right bundle branch block, AUROC of 0.79 (95% CI 0.77- 0.81) and AUPRC of 0.18 (95% CI 0.15- 0.21) were obtained.

[0539] Survival analysisPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0540] FIG. 46 illustrates Kaplan-Meier survival analysis based on artificial intelligence- enhanced electrocardiogram risk stratification. Kaplan-Meier curve survival analysis when stratifying patients as low- (blue) or high-risk (orange) based on artificial intelligence-enhanced electrocardiogram predictions using random (left), last (middle), or first (right) electrocardiogram per patient. Number at risk within each group inset below. Hazard ratio inset below (high- vs. low-risk group based on artificial intelligence-enhanced electrocardiogram predictions) with 95% confidence interval using Cox regression analysis. P-value statistic below based on log-rank testing. ECG, electrocardiogram; HR, hazard ratio.

[0541] Longer-term survival was assessed when stratifying patients into low- (<threshold) or high-risk (^threshold) groups based on ALECG predictions (FIG. 46). When using a random ECG, there was 15-year survival of 96% and 80% for low- vs. high-risk ECGs, respectively. When using a random ECG, high-risk patients were 4.9 times (95% CI 4.3-5.6) more likely to experience mortality (P < .001). This phenomenon was more pronounced when using the last ECG, with 15-year survival of 95% and 72% in low- and high-risk groups, respectively, and a hazard ratio of 7.2 (95% CI 6.3-8.3) (P < .001).

[0542] In the internal test cohort, the AI-ECG model predictions [c-index 0.74 (95% CI 0.72-0.76)] outperformed QRS duration [c-index 0.52 (95% CI 0.50-0.54)] and LVEF [c-index 0.64 (95% CI 0.60-0.68)] in Cox model survival discrimination (TABLE 20). The addition of age and / or LVEF as predictors to AI-ECG provided no added value in the c-index (TABLE 20).

[0543] TABLE 20 Cox models on the internal test cohort incorporating artificial intelligence-enhanced electrocardiogram predictionsPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0544] AI-ECG, artificial intelligence-enhanced electrocardiogram; CI, confidence interval; LVEF, left ventricular ejection fraction.

[0545] FIG. 47 illustrates Lesion-specific Kaplan-Meier survival analysis based on artificial intelligence-enhanced electrocardiogram risk stratification. Lesion-specific Kaplan- Meier curve survival analysis when stratifying patients as low- (blue) or high-risk (orange) based on artificial intelligence-enhanced electrocardiogram predictions. Hazard ratio inset below (high- vs. low-risk group based on artificial intelligence-enhanced electrocardiogram predictions) with 95% confidence interval using Cox regression analysis. P-value statistic below based on log-rank testing. Initial sample size in each cohort inset. ASD, atrial septal defect; CoA, coarctation of the aorta; DORV, double outlet right ventricle; HR, hazard ratio; HLHS, hypoplastic left heart syndrome; PA, pulmonary atresia; ToF, tetralogy of Fallot; TAPVC, total anomalous pulmonary venous connection; TGA, transposition of the great arteries; VSD, ventricular septal defect.

[0546] Lesion-specific survival analyses demonstrated effective risk stratification in nearly all lesions (FIG. 47), including dextrocardia [hazard ratio 2.4 (95% CI 1.1-5.3); P = .02; data not provided), but not complete atrioventricular canal defects [hazard ratio 3.5 (95% CI 0.8- 15.8); P = .10; data not provided],

[0547] Model explainability

[0548] FIG. 48 illustrates explainability of artificial intelligence-enhanced electrocardiogram predictions. Visualization of median waveforms generated in each lead using electrocardiograms from the highest (red) and lowest (green) artificial intelligence-enhanced electrocardiogram predictions of the overall cohort, as well as cardiomyopathy, tetralogy of Fallot, and hypoplastic left heart syndrome subgroups. Saliency mapping demarcates regions of the electrocardiogram waveform having greatest (dark blue) and least (light blue) influence on each outcome. Saliency was averaged over the highest predicted electrocardiograms for each outcome. HLHS, hypoplastic left heart syndrome; ToF, tetralog}7of Fallot.

[0549] The most salient features (FIG. 48) of an ECG to predict 5-year mortality in the overall cohort include S waves (limb lead III and precordial leads V2, V3, and V6) and T waves (limb leads II and V6). High-risk features to predict 5-year mortality include wide QRS complexes with deep S waves and low-amplitude waveforms (FIG. 48).

[0550] When performing saliency mapping and median waveform analysis within subgroups, lesion-specific high-risk signatures were identified (Figure 6). In cardiomyopathy, V2 and V3 were less salient than the overall cohort; however, high-risk features appeared similar to the overall cohort. In right-sided myopathy such as ToF, V6 was less salient than the overall cohort; in addition, while low-risk features included a right bundle branch block, high-riskPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 features included QRS fragmentation. In left-sided pathology such as hypoplastic left heart syndrome, saliency maps were more focused on the QRS complex compared with the overall cohort. High-risk features included tall and wide R waves in V 1-3 with a deep S wave in II— III and lateral precordial leads.

[0551] DISCUSSION

[0552] FIG. 49 provides a central illustration of a large and diverse pediatric and adult congenital heart disease cohort used to train and test an artificial intelligence-enhanced electrocardiogram algorithm to accurately predict 5-year mortality across a range of congenital heart disease lesions. In an effort to interpret model behavior, model explainability analysis was performed. AI-ECG, artificial intelligence-enhanced electrocardiogram; ASD, atrial septal defect; CAVC, complete atrioventricular canal defect; CNN, convoluted neural network; CoA, coarctation of the aorta; DORV, double outlet right ventricle; ECG, electrocardiogram; HLHS, hypoplastic left heart syndrome; LV, left ventricle; LVEF, left ventricular ejection fraction; PA, pulmonary atresia; RV, right ventricle; TAPVR, totally anomalous pulmonary venous return; TGA, transposition of the great arteries; ToF, tetralogy of Fallot; VSD, ventricular septal defect.

[0553] Risk stratification has been of great interest in the CHD field using conventional and artificial intelligence approaches. The ongoing challenge in developing robust risk prediction models in CHD has led to multiple recent calls for Al applications to improve risk stratification in CHD. In this work, this gap was addressed by developing and validating an ECG-based deep learning algorithm to predict mortality in children and adults with CHD. AI-ECG was demonstrated to successfully predict 5-year mortality, outperforming conventional markers such as age, LVEF, and QRS duration. The encouraging performance when using the first available, the last available, or a random ECG per patient demonstrates the promise of the model to predict mortality during initial and follow- up assessments to aid in lifelong risk stratification. Model explainability analysis provides transparency and interpretability' for clinicians and may help generate hypotheses for underlying myopathy ECG signatures predictive of mortality.Altogether, these findings demonstrate the prognostic value of AI-ECG in CHD across the lifespan, which may (i) enhance current risk stratification strategies, (ii) prioritize patients for diagnostic studies and / or interventions, and (iii) facilitate improved access to care (FIG. 49).

[0554] Conventional electrocardiogram predictors of mortality in congenital heart disease

[0555] There is a paucity of conventional ECG analyses to predict morbidity and mortality in patients with CHD. The majority of applications are in ToF: several studies have recognized severe QRS prolongation as a risk factor for mortality in this patient population. However, the sensitivity of QRS duration > 180 ms to predict mortality was <50% in re- centPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 studies. Subsequent work has demonstrated QRS fragmentation — related to myocardial fibrosis and dysfunction — is superior to QRS duration in predicting mortality. QRS fragmentation has also been associated with mortality in cardiomyopathy. In addition, a longer QRS has been previously predictive of mortality in hypoplastic left heart sy ndrome, D-loop transposition of the great arteries, and Fontan circulation, all with limited performance. This has motivated more recent efforts to incorporate imaging modal- ity data into risk prediction algorithms. However, even such algorithms continue to have suboptimal performance and are reliant on expensive modalities (e.g. cardiac magnetic resonance imaging) that require subspecialized expertise. As shown herein, AI-ECG provides an inexpensive, ubiquitous alternative that is predictive across a wide range of lesions.

[0556] Clinical significance and implications

[0557] Imaging modalities (e.g. cardiac magnetic resonance and echocardiography) conventionally used to aid in risk stratification in CHD have practice limitations (time-intensive, resource-consuming, and need for subspecialist expertise) that hinder its widespread use. In contrast, ECGs is rapid and cost-effective and can be conveniently acquired at every cardiology visit, which facilitates easier and more frequent use that may guide clinical decision-making. This AI-ECG algorithm may serve as a screening or surveillance tool and potentially improve access to care.

[0558] Given the objective to risk stratify patients from a preventative lens, training and testing used ECGs from cardiology clinic. In this setting, AI-ECGs could be of significant screening and surveillance value. A NPV of 98% was achieved to predict 5-year mortality, with 15-year survival of 96% for low-risk ECGs (data not provided). To this end, as a screening tool, low-risk AI-ECG predictions may help reduce follow-up frequency, the need for non-invasive imaging, diagnostic catheterizations, or implantable cardioverter-defibrillators. On the other hand, given the 5-year mortality PPV of 12%, it may help identify high-risk patients requiring closer monitoring.

[0559] As a surveillance tool, a congenital cardiologist could conceivably monitor AI- ECG predictions at each cardiology visit. Monitoring AI-ECG predictions over time may provide insight into responsiveness to interventions (e.g. pulmonary valve replacement in ToF) or interstage monitoring of single ventricle patients.

[0560] Finally, for low-resource settings with limited access to advanced modalities, this algorithm may help improve access to care. Notably, despite the majority of CHD patients being adults, approximately half remain without regional CHD services. This technology may therefore contribute to the democratization of specialty expertise and circumvent the requirement for specialized cardiac magnetic resonance knowledge to risk stratify certain lesions (e.g. ToF).PCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025

[0561] Significance of temporal validation

[0562] From 1990 (the first ECG available in this cohort) through today, CHD has rapidly evolved. For example, surgical and medical advances have contributed to the nearly 40% decrease in CHD mortality from 1999 to 2017,46 with >97% of children with CHD expected to reach adulthood. From an institutional standpoint, the Boston Children’s Hospital volume and complexity have significantly increased (TABLE 20). Given the range of confounding factors that could affect mortality, temporal validation was performed with reassuringly similar performance.

[0563] Artificial intelligence-enhanced electrocardiogram model insights gained

[0564] The overall CHD and cardiomyopathy cohorts had similar high-risk features, suggesting common final signatures on an ECG indicative of high-risk mortality. Indeed, both CHD and cardiomyopathy have non-trivial overlap, and both largely contribute to pediatric heart failure. The common features noted herein include wide QRS complexes with deep S waves and low-amplitude waveforms, which may correspond to heterogeneous slow activation of the myocardium due to scar and / or myocardial stress. In addition, select high-risk patterns were noted in each disease process. For example, QRS fragmentation was noted in high-risk ECGs for ToF (Figure 6). Future work is required to investigate the relation of these features with progressive cardiomyopathy and pediatric heart failure.

[0565] Finally , variation in model performance was noted by disease subtype. Interestingly, effective risk stratification was achieved in a majority of diseases (including dextrocardia), but not complete atrioventricular canal defects. The reasons for poor performance in this group are not readily apparent but may involve the high heterogeneity of this group given its association with heterotaxy syndrome and situs abnormalities, trisomy 21, and inherently abnormal QRS axis (superior axis deviation) compared with the rest of the cohort.

[0566] Limitations and future directions

[0567] There are several limitations of this work. First, although heart failure is the most common cause of death in adult CHD, all-cause mortality rather than cardiac mortality was used for the primary outcome in this study. In addition, while all-cause mortality is routinely coded internally, it is conceivable that positive outcomes are undocumented. This limitation was mitigated in binary outcome analysis by including only ECGs with mortality events within the outcome timeframe or documented follow-up after the outcome timeframe. Similarly, in survival analysis, censoring was performed at time of the last known follow-up. Nevertheless, future use of state or national death index database should be considered. Second, while several recent AI- ECG works also used all-cause mortality as the primary endpoint, similar clinically meaningful outcomes such as heart transplant are also of interest. Third, while temporal validation wasPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 achieved, it is of great interest to obtain external validation for each CHD lesion. Fourth, only one example of thresholding was used to evaluate model performance, as further consideration is required to weigh the impact of resultant false negatives (which may lead to clinical consequences of missed pathology) and false positives (which may lead to un- necessary extraneous testing), as well as optimally set thresholds across institutions. To this end, multicenter external validation to further refine thresholds for clinical implementation is warranted. Similarly, multicenter collaboration via federated learning may help improve training / testing sample sizes, which may further im- prove performance. Given the objective to develop an inexpensive and convenient risk stratification tool, only ECG inputs were utilized; nevertheless, multimodal inputs may lead to improved model performance (especially for complex lesions) requiring further investigation. The limitations of saliency mapping must be noted. Lastly, diagnostic categories in this study are quite heterogeneous. For example, the cardiomyopathy category includes dilated, hypertrophic, and restrictive cardiomyopathy, among others. Similarly, multiple CHD diagnoses can occur simultaneously, such that patients can be assigned into multiple categories. In addition, the cohort includes patients with and without repairs for CHD lesions.

[0568] Future work therefore includes model refinement and external validation for each lesion of interest, multicenter collaboration, consideration of multimodal inputs, and prospective trials (to determine how to properly implement such tools to support clinical decision-making). Finally, a recent randomized clinical trial implementing AI-ECG alerts led to decreased all-cause mortality in the general adult population. It is similarly of great interest to implement a similar study design in the distinct CHD population.

[0569] CONCLUSIONS

[0570] In conclusion, these findings demonstrate the promise of AI-ECG to inexpensively and conveniently risk stratify individuals with CHD across the lifespan. This tool may facilitate the prioritization of patients for future interventions / studies, provide meaningful insight into novel ECG waveforms suggestive of mortality, and potentially reduce disparities by improving access to care. Future multicenter collaboration and prospective trials are warranted.

[0571] Embodiments have been described where the techniques are implemented in circuitry and / or computer-executable instructions. It should be appreciated that some embodiments may be in the form of a method, of which at least one example has been provided. The acts performed as part of the method may be ordered in any suitable way. Accordingly, embodiments may be constructed in which acts are performed in an order different than illustrated, which mayPCT / US25 / 48621 30 September 2025 (30.09.2025)Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 include performing some acts simultaneously, even though shown as sequential acts in illustrative embodiments.

[0572] Various aspects of the embodiments described above may be used alone, in combination, or in a variety' of arrangements not specifically discussed in the embodiments described in the foregoing and is therefore not limited in its application to the details and arrangement of components set forth in the foregoing description or illustrated in the drawings. For example, aspects described in one embodiment may be combined in any manner with aspects described in other embodiments.

[0573] Use of ordinal terms such as “first,” “second,” “third,” etc., in the claims to modify a claim element does not by itself connote any priority, precedence, or order of one claim element over another or the temporal order in which acts of a method are performed, but are used merely as labels to distinguish one claim element having a certain name from another element having a same name (but for use of the ordinal term) to distinguish the claim elements.

[0574] Also, the phraseology and terminology used herein is for the purpose of description and should not be regarded as limiting. The use of “including,” “comprising,” “having,” “containing,” “involving,” and variations thereof herein, is meant to encompass the items listed thereafter and equivalents thereof as well as additional items.

[0575] The word “exemplary” is used herein to mean serving as an example, instance, or illustration. Any embodiment, implementation, process, feature, etc. described herein as exemplary' should therefore be understood to be an illustrative example and should not be understood to be a preferred or advantageous example unless otherwise indicated.

[0576] Having thus described several aspects of at least one embodiment, it is to be appreciated that various alterations, modifications, and improvements will readily occur to those skilled in the art. Such alterations, modifications, and improvements are intended to be part of this disclosure, and are intended to be within the spirit and scope of the principles described herein. Accordingly, the foregoing description and drawings are by way of example only.

Claims

Attorney Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025CLAIMSWhat is claimed is:

1. A method of predicting a cardiac abnormality in a pediatric patient or adult congenital patient, the method comprising: determining, based on received electrocardiogram (ECG) data of the pediatric patient or adult congenital patient, one or more of a plurality of cardiac phenotypes to which to assign the pediatric patient or adult congenital patient, at least one cardiac phenotype of the plurality of cardiac phenotypes each corresponding to presence of a cardiac abnormality, the determining the one or more of the plurality of cardiac phenoty pes to assign to the pediatric patient or adult congenital patient comprising analyzing the received ECG data using one or more trained models, wherein the one or more trained models were trained with training data comprising ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert; and outputting the one or more of the plurality of cardiac phenotypes determined for the pediatric patient or adult congenital patient based on the ECG data.

2. The method of claim 1 , wherein determining the one or more of a plurality of cardiac phenotypes to which to assign the pediatric patient or adult congenital patient comprises determining a probability of the pediatric patient or adult congenital patient having one or more cardiac abnormalities.

3. The method of claim 1, wherein: determining a probability of the pediatric patient or adult congenital patient having one or more cardiac abnormalities comprises determining a probability of the pediatric patient or adult congenital patient having any of the one or more cardiac abnormalities; and determining the one or more of the plurality of cardiac phenotypes to which to assign the pediatric patient or adult congenital patient comprises, based on the probability of the pediatric patient or adult congenital patient having any of the one or more cardiac abnormalities, determining whether to assign the pediatric patient or adult congenital patient to a phenotype for pediatric patient or adult congenital patients with any of the one or more cardiac abnormalities.Attomey Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 20254. The method of claim 2, wherein determining a probability of the pediatric patient or adult congenital patient having one or more cardiac abnormalities comprises determining a probability’ of the pediatric patient or adult congenital patient having each of the one or more cardiac abnormalities, wherein the one or more cardiac abnormalities comprise mortality7, ventricular dysfunction, ventricular dilation, ventricular hypertrophy, one or more rhythm abnormality , or critical congenital heart disease.

5. The method of claim 4, wherein determining the one or more of the plurality of cardiac phenoty pes comprises comparing the probability7of the pediatric patient or adult congenital patient having each of the one or more cardiac abnormalities to a threshold.

6. The method of claim 1 , wherein determining one or more of the plurality of cardiac phenotypes comprises determining one or more of a mortality7risk, ventricular dysfunction phenoty pe, ventricular dilation phenotype, ventricular hypertrophy phenoty pe, one or more rhythm abnormality, or critical congenital heart disease.

7. The method of claim 1, wherein determining one or more of the plurality of cardiac phenoty pes comprises determining whether the pediatric patient or adult congenital patient has a ventricular dysfunction.

8. The method of claim 1, wherein determining one or more of the plurality of cardiac phenoty pes comprises determining whether the pediatric patient or adult congenital patient has ventricular dilation.

9. The method of claim 1, wherein determining one or more of the plurality of cardiac phenoty pes comprises determining whether the pediatric patient or adult congenital patient has ventricular hypertrophy.

10. The method of claim 1, wherein determining one or more of the plurality of cardiac phenotypes comprises determining whether the pediatric patient or adult congenital patient has a risk of mortality.

11. The method of claim 1 , wherein determining one or more of the plurality of cardiac phenotypes further comprises assigning a qualitative assessment for the one or more of theAttomey Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 plurality of cardiac phenotypes, wherein the qualitative assessment indicates whether a severity of a cardiac abnormality indicated by a cardiac phenotype.

12. The method of claim 1, wherein determining the one or more of the plurality of cardiac phenoty pes based on the received ECG data comprises determining the one or more of the plurality of cardiac phenotypes based on one or more of: at least one ECG raw waveform, a QRS interval, a QRS axis, a T axis, a P axis, a PR interval, a QT interval, a QT corrected for heart rate (QTc), or a heart rate.

13. The method of claim 1, wherein determining using the one or more trained models comprises determining using the one or more trained models were trained with training data that did not include data for pediatric patients having congenital heart disease.

14. The method of claim 1, wherein determining using the one or more trained models comprises determining using the one or more trained models were trained with training data including data for pediatric patient or adult patients having congenital heart disease.

15. The method of claim 1, wherein determining using the one or more trained models comprises determining using one or more trained models were trained with training data for patients across a plurality of age ranges, each of the plurality of age ranges being a range within an overall age range of 0 to 18 years, and greater than 18 years for adult congenital heart disease.

16. The method of claim 15, wherein the plurality of age ranges comprise two or more of: 0 years to 1 year. 1 year to 3 years, 3 years to 8 years. 8 years to 12 years, 12 years to 18 years, or greater than 18 years in adult congenital heart disease.

17. The method of claim 15, wherein the plurality of age ranges comprise two or more of: infancy, toddlerhood, school age, preadolescence, adolescence, and adults.

18. The method of claim 1, wherein determining using the one or more trained models comprises determining using one or more trained models were trained with training data comprising age data and / or sex data from the plurality of prior pediatric patient or adult congenital patients, wherein each of the age data and / or the sex data is combined in the training with the ECG data from the plurality’ of prior pediatric patient or adult congenital patients.Attomey Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 202519. The method of claim 1, wherein determining using the one or more trained models comprises determining using one or more trained models were trained with training data comprising echocardiogram data from the plurality of prior pediatric patient or adult congenital patients, wherein the echocardiogram data is combined with the ECG data from the plurality7of prior pediatric patient or adult congenital patients.

20. The method of claim 19, wherein the echocardiogram data of the training data comprises one or more of a death date, ventricular ejection fraction EF, a ventricular mass, a ventricular mass / volume, and a ventricular end-diastolic volume.

21. The method of claim 1, wherein determining using one or more trained models trained with ECG data from the plurality of prior pediatric patient or adult congenital patients comprises determining using one or more trained models trained with filtered ECG data for the plurality7of prior pediatric patient or adult congenital patients, the filtered ECG data comprising noise exceeding a threshold.

22. A method of predicting a cardiac abnormality in a pediatric patient or adult congenital patient, the method comprising: determining, based on received ECG data from a pediatric patient or adult congenital patient, a prediction of whether the pediatric patient or adult congenital patient has any of one or more cardiac abnormalities, the determining the prediction comprising analyzing the received ECG data using one or more trained models, wherein the one or more trained models were trained with training data comprising ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert; and outputting the prediction of whether the pediatric patient or adult congenital patient has any of the one or more cardiac abnormalities.

23. A method of predicting cardiac abnormalities in pediatric patients or adult congenital patients, the method comprising: determining, based on first received ECG data from a first pediatric patient or adult congenital patient, a first prediction of whether the first pediatric patient or adult congenital patient has one or more cardiac abnormalities, the determining comprising analyzing the first received ECG data using a trained model;Attomey Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 determining, based on second received ECG data from a second pediatric patient or adult congenital patient, a second prediction of whether the second pediatric patient or adult congenital patient has the one or more cardiac abnormalities, the determining comprising analyzing the second received ECG data using the trained model; and outputting the first prediction and the second prediction for the first pediatric patient or adult congenital patient and the second pediatric patient or adult congenital patient, wherein the trained model was trained with training data comprising ECG data and cardiac abnormality data from a plurality of prior pediatric patient or adult congenital patients comprising patients from multiple age ranges of pediatric patient or adult congenital patients, the multiple age ranges comprising at least one adolescent age range and at least one preadolescent age range, and wherein the first pediatric patient or adult congenital patient has an age in the at least one adolescent age range and the second pediatric patient or adult congenital patient has an age in the at least one preadolescent age range, and wherein the one or more trained models were trained with training data comprising ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert.

24. A system for predicting a cardiac abnormality in a pediatric patient or adult congenital patient, the system comprising: an ECG monitoring device; a controller comprising at least one processor, at least one storage medium having encoded thereon executable instructions that, when executed by the at least one processor, cause the at least one processor to carry out a method comprising: determining, based on received ECG data from a pediatric patient or adult congenital patient, information regarding whether the pediatric patient or adult congenital patient has one or more cardiac abnormalities, the determining comprising analyzing the received ECG data using one or more trained model, and wherein the one or more trained models were trained with training data comprising ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert; andAttomey Docket No.: 167705-037401 / PCT Electronic Deposit Date: September 30, 2025 outputting the information regarding whether the pediatric patient or adult congenital patient has the one or more cardiac abnormalities.

25. At least one storage medium having encoded thereon executable instructions that, when executed by at least one processor, cause the at least one processor to carry out a method comprising: determining, based on received ECG data from a pediatric patient or adult congenital patient, information regarding whether the pediatric patient or adult congenital patient has one or more cardiac abnormalities, the determining comprising analyzing the received ECG data using one or more trained model, and wherein the one or more trained models were trained with training data comprising ECG data from a plurality of prior pediatric patients or adult congenital patients and information indicating whether each prior patient of the plurality of prior pediatric patient or adult congenital patients had one or more cardiac abnormalities based on a readjudication from an expert; and outputting the information regarding whether the pediatric patient or adult congenital patient has the one or more cardiac abnormalities.

Images