Determining presence of structural changes in the heart from ECG using deep learning

Abstract

Machine learning techniques for assessing cardiac remodeling in a patient based on electrocardiogram (ECG) data are presented. The techniques include: acquiring patient data including patient electrocardiogram (ECG) data; passing the patient data as an input to a trained machine learning model, where the trained machine learning model has been trained with a corpus of training data that includes labels, where the corpus of training data includes ECG data, and where the labels are indicative of a presence of cardiac remodeling; obtaining a patient cardiac remodeling indication output from the trained machine learning model, where the patient cardiac remodeling indication is indicative of whether the patient has cardiac remodeling; and providing an assessment of cardiac remodeling in the patient based on the cardiac remodeling indication.

Description

DETERMINING PRESENCE OF STRUCTURAL CHANGES IN THE HEART FROM ECG USING DEEP LEARNINGRelated Application

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 569,284, entitled “Determining Presence of Structural Changes in the Heart From ECG Using Deep Learning,” and filed 25 March 2025.Government Funding

[0002] This invention was made with government support under grant HL166759 awarded by the National Institutes of Health. The government has certain rights in the invention.Field

[0003] This disclosure relates generally to cardiac substrate remodeling diagnostics.Background

[0004] Atrial fibrillation (AF), the most common cardiac arrhythmia globally, significantly compromises patient quality of life. AF is driven by various pathophysiological mechanisms. This condition leads to an increase in hospitalizations and mortality rates, exerting a considerable financial burden on healthcare systems due to the extensive costs associated with its management. Notably, atrial substrate remodeling, e.g., atrial fibrosis, has emerged as a pivotal factor in the pathogenesis of AF.

[0005] Ventricular arrhythmias (VA) are the main cause of sudden cardiac deaths, the latter being the largest cause of mortality worldwide. Disease-induced substrate remodeling in the ventricles leading to arrhythmias may entail scarring, fibrosis, inflammation, change in neural innervation, etc.

[0006] Thus, substrate remodeling in the atria or ventricles of the heart, e.g., changes in the tissues of the heart caused by disease, lifestyle, and other factors, can lead to a variety of adverse outcomes in patients, including sudden death and failed treatment (e.g., ablation). Studies have underscored the critical role of prompt detection and intervention. This evidence highlights the paramount importance of early and accurate identification of cardiac substrate remodeling for personalized risk stratification and arrhythmia treatment planning, ultimately aiming to enhance patient outcomes.

[0007] The most accurate technique for cardiac substrate evaluation of the ventricles is late gadolinium enhancement magnetic resonance imaging (LGE MRI). LGE-MRI is non-invasive and provides a visualization of the most common remodeled substrate (scarring or fibrosis) within the ventricular anatomy. LGE-MRI is typically acquired for patients undergoing ventricular arrhythmia risk assessment so that measures to prevent sudden cardiac death can be undertaken. However, LGE-MRI is very expensive and typically not prescribed to every patient that could be at risk. Further, the practicality of deploying this comprehensive imaging modality for every patient, especially due to limitations of availability, time, resource demands, and occasionally subpar image quality, is challenging. Thus, although LGE MRI is effective as diagnostic tool for scar / fibrosis-dependent arrhythmias, it encounters challenges in widespread application, particularly in primary and secondary healthcare centers.

[0008] Visualization of remodeling is used in different ways in ventricular and atrial arrhythmia management. In ventricular arrhythmias, which can be lethal, it is used to assess risk of sudden cardiac death and guide decisions regarding prophylactic implantation of defibrillation devices for primary prevention. It can also be used in the treatment of VA (e.g., ablation).

[0009] By contrast, in atrial arrhythmia management, where the arrhythmias are generally not lethal, visualization of remodeling is helpful in the treatment of atrial arrhythmia via ablation, and also in the assessment of the risk of stroke, as the presence of remodeling can lead to tromboembolic events. However, visualizing scar and fibrosis by LGE-MRI in the atria is difficult because of the thin atrial walls, and it is not used routinely in clinical practice. Accordingly, in the atria, low voltage electroanatom ical mapping (EAM) is frequently used as a surrogate for LGE-MRI. However, because acquisition of EAM is invasive, it can only be done during an atrial ablation procedure.

[0010] Electrocardiograms (ECGs) are a non-invasive, inexpensive, and widely accessible diagnostic tool. Typically, ECGs record cardiac voltage using a standard array of 12 external electrodes positioned on the patient.Summary

[0011] According to various embodiments, a machine learning method of assessing cardiac remodeling in a patient is presented. The method includes: acquiring patient data including patient electrocardiogram (ECG) data; passing the patient data as an input to a trained machine learning model, where the trained machine learning model has been trained with a corpus of training data including labels, where the corpus of training data includes ECG data, and where the labels areindicative of a presence of cardiac remodeling; obtaining a patient cardiac remodeling indication output from the trained machine learning model, where the patient cardiac remodeling indication is indicative of whether the patient has cardiac remodeling; and providing an assessment of cardiac remodeling in the patient based on the cardiac remodeling indication.

[0012] Various optional features of the above method embodiments include the following. The labels may include labels representative of a presence of cardiac remodeling in atria, and the patient cardiac remodeling indication may be indicative of atrial cardiac remodeling in the patient. The labels may include labels representative of a presence of low voltage areas in atria, and the patient cardiac remodeling indication may be indicative of a low voltage area in a cardiac atrium of the patient. The labels may include labels representative of locations of low voltage areas in atria, and the patient cardiac remodeling indication may be indicative of a location of the low voltage area in the cardiac atrium of the patient. The labels may include labels representative of atrial fibrillation classification, and the patient cardiac remodeling indication may be indicative of an atrial fibrillation classification for the patient. The method may include providing a prognosis of the patient based on the assessment of cardiac remodeling. The method may include providing a treatment to the patient based on the assessment of cardiac remodeling, where the treatment includes performing cardiac ablation on the patient. The method may include determining a location of the cardiac ablation based on the assessment of cardiac remodeling. The method may include determining a type of the cardiac ablation based on the assessment of cardiac remodeling. The labels may include labels representative of a presence of cardiac remodeling in ventricles, and the patient cardiac remodeling indication may be indicative of ventricle cardiac remodeling in the patient. The methodmay include, based on the assessment of cardiac remodeling: imaging a cardiac ventricle of the patient using at least one of: late gadolinium enhanced magnetic resonance imaging (LGE MRI), echocardiogram, or positron emission tomography (PET); and predicting heart disfunction based on the imaging. The method may include providing a treatment or prophylaxis to the patient based on the predicting, where the treatment or prophylaxis includes at least one of: cardiac ablation, anticoagulant administration, or defibrillation implantation. The patient data may further include demographic data for the patient, and the corpus of training data may further include demographic data of individuals. The trained machine learning model may include a convolutional neural network, a long short-term memory network, and a cross-attention model. The method may include performing a longitudinal study of the patient by repeating the acquiring, the passing, the obtaining, and the providing over a period of time.

[0013] According to various embodiments, a machine learning system for assessing cardiac remodeling in a patient is presented. The system includes a non- transitory computer readable medium including instructions, and at least one electronic processor that executes the instructions to perform operations including: acquiring patient data including patient electrocardiogram (ECG) data; passing the patient data as an input to a trained machine learning model, where the trained machine learning model has been trained with a corpus of training data including labels, where the corpus of training data includes ECG data, and where the labels are indicative of a presence of cardiac remodeling; obtaining a patient cardiac remodeling indication output from the trained machine learning model, where the patient cardiac remodeling indication is indicative of whether the patient has cardiac remodeling; andproviding an assessment of cardiac remodeling in the patient based on the cardiac remodeling indication.

[0014] Various optional features of the above system embodiments include the following. The labels may include labels representative of a presence of cardiac remodeling in atria, and the patient cardiac remodeling indication may be indicative of atrial cardiac remodeling in the patient. The labels may include labels representative of a presence of low voltage areas in atria, and the patient cardiac remodeling indication may be indicative of a low voltage area in a cardiac atrium of the patient. The labels may include labels representative of locations of low voltage areas in atria, and the patient cardiac remodeling indication may be indicative of a location of the low voltage area in the cardiac atrium of the patient. The labels may include labels representative of atrial fibrillation classification, and the patient cardiac remodeling indication may be indicative of an atrial fibrillation classification for the patient. The labels may include labels representative of a presence of cardiac remodeling in ventricles, and the patient cardiac remodeling indication may be indicative of ventricle cardiac remodeling in the patient. The patient data may further include demographic data for the patient, and the corpus of training data may further include demographic data of individuals. The trained machine learning model may include a convolutional neural network, a long short-term memory network, and a cross-attention model.

[0015] Combinations, (including multiple dependent combinations) of the above-described elements and those within the specification have been contemplated by the inventors and may be made, except where otherwise indicated or where contradictory.Brief Description of the Drawings

[0016] Various features of the examples can be more fully appreciated, as the same become better understood with reference to the following detailed description of the examples when considered in connection with the accompanying figures, in which:

[0017] Fig. 1 illustrates a cardiac disease progression according to various embodiments;

[0018] Fig. 2 is a schematic diagram of a machine learning system for assessing cardiac atrial substrate remodeling in a patient using ECG data, according to various embodiments;

[0019] Fig. 3 illustrates 3D electroanatom ical maps (EAMs) acquired from the left atrium before and after excluding the areas of pulmonary veins, mitral valve, and left atrial appendage, generating the low voltage map that serves as a surrogate for atrial remodeling, according to various embodiments;

[0020] Fig. 4 is a schematic diagram of a machine learning system for assessing cardiac ventricular substrate remodeling in a patient according to various embodiments;

[0021] Fig. 5 is a flow diagram of a machine learning method of assessing cardiac remodeling in a patient according to various embodiments;

[0022] Fig. 6 is a flow diagram of method of following up an assessment of atrial cardiac remodeling by the method of Fig. 5, according to various embodiments; and

[0023] Fig. 7 is a flow diagram of method of following up an assessment of ventricle cardiac remodeling by the method of Fig. 5, according to various embodiments.Description of the Examples

[0024] Reference will now be made in detail to example implementations, illustrated in the accompanying drawings. Wherever convenient, the same reference numbers will be used throughout the drawings to refer to the same or like parts. In the following description, reference is made to the accompanying drawings that form a part thereof, and in which is shown by way of illustration specific exemplary examples in which the invention may be practiced. These examples are described in sufficient detail to enable those skilled in the art to practice the invention and it is to be understood that other examples may be utilized and that changes may be made without departing from the scope of the invention. The following description is, therefore, merely exemplary.

[0025] There is a need for non-invasive, efficient, economical, and accessible cardiac substrate remodeling diagnostic tools. Some embodiments fulfill this need through the use of patient ECG data together with machine learning as a tool for assessing cardiac substrate remodeling, thereby facilitating the development of more informed, effective, and personalized patient follow-up and treatment strategies in clinical settings.

[0026] For example, some embodiments employ deep learning techniques to exploit the conventional 12-lead ECG for predicting cardiac substrate remodeling in cardiac atria and ventricles. For cardiac remodeling of atrial substrate, some embodiments predict one or more of: the presence of cardiac atria substrate remodeling, the existence of atrial low voltage areas (LVAs), the extent of atrial LVAs, the location of atrial LVAs, and / or the classification of AF (e.g., persistent or paroxysmal AF). For cardiac remodeling of ventricle substrate, some embodimentspredict the presence of cardiac ventricle substrate remodeling and / or the extent of cardiac ventricle substrate remodeling.

[0027] Embodiments may inform various follow-up and treatment strategies for patients. By way of non-limiting example, the outputs of a trained machine learning model may be used to determine an appropriate treatment strategy for a patient for whom atrial cardiac substrate remodeling is predicted. Such treatment strategies may include one or more of: whether to perform ablation, an ablation location, and / or an ablation type. As another non-limiting example, the outputs of a trained machine learning model may be used to determine an appropriate follow-up strategy for a patient for whom ventricle cardiac substrate remodeling is predicted. Such follow-up strategies may include: whether to perform comprehensive cardiac imaging (e.g., LGE MRI, echocardiogram, and / or positron emission tomography (PET)), and / or whether to proceed with treatment (e.g., ablation, anticoagulant administration) or prophylaxis (e.g., an implantable cardioverter-defibrillator (ICD)).

[0028] Thus, some embodiments provide enhanced early diagnosis and risk stratification. For example, by leveraging Al in ECG analysis, some embodiments may significantly improve the early diagnosis of comorbidities in AF patients. Early and accurate diagnosis may allow for better risk stratification and management of these patients. Some embodiments move beyond the limitations of current cardiac diagnostic methods, many of which are invasive, resource-intensive, or lack granularity in early-stage disease detection. Some embodiments provide personalized treatment strategies. For example, by accurately identifying and monitoring comorbidities and the progression of atrial and cardiac remodeling, healthcare providers can tailor treatment plans more effectively, potentially improving success rates of interventions like pulmonary vein isolation (PVI) and reducing recurrence ofAF. Some embodiments streamline the diagnostic process, making it more cost- effective and accessible. This is particularly beneficial in settings with limited access to advanced diagnostic tools, thereby addressing a significant gap in global cardiac care. Thus, some embodiments provide improved diagnostic and treatment approaches for cardiac patients (e.g., AF patients), which can lead to better overall health outcomes, reduced hospitalizations, and a decrease in the healthcare burden associated with cardiac diseases.

[0029] These and other features and advantages are shown and described herein in reference to the accompanying figures.

[0030] Fig. 1 illustrates a cardiac disease progression 100 according to various embodiments. In particular, Fig. 1 illustrates a typical progression from inflammation (or other primary cause, such scarring or fibrosis, e.g., from cardiac sarcoidosis), to myocardial scarring, and then on to ventricular arrhythmia (VA) and, potentially, sudden cardiac death. Some embodiments may be used to arrest this progression by detecting ventricular substrate remodeling early, e.g., before VA is present. Upon detection of ventricular substrate remodeling by such embodiments, various imaging may be utilized (e.g., LGE MRI, echocardiogram, or PET) to determine whether, and / or what type, of any of a variety of treatments (e.g., ablation, anticoagulant administration) or prophylaxis (e.g., ICD implantation) may be implemented. For example, embodiments may utilize the remodeling assessment to determine whether a particular patient would benefit from prophylactic implantation of an ICD.

[0031] Note that embodiments are not limited to detection of cardiac substrate remodeling in cardiac ventricles. Some embodiments may be used to detect cardiac substrate remodeling in cardiac atria. Upon detection of atrial substrate remodeling by such embodiments, various treatments may be implemented, e.g., PVI, potentiallywith ablation at one or more other locations. Some embodiments may inform one or more locations of cardiac ablation, e.g., by predicting a location of LVA. Once such location(s) are provided according to some embodiments, a type of ablation may be determined. For example, if PVI alone is to be performed, then cryoablation can be done, which is fast, although it generally cannot be used to ablate away from the pulmonary veins. By contrast, if PVI is performed along with ablation at a different LVA, then a different type of ablation, e.g., radiofrequency ablation, may be performed. Note that embodiments are not limited to use with cryoablation and radiofrequency ablation. Another suitable type of ablation is pulse field ablation (PFA), which can be used for both PVI and well as ablation outside of PVI.

[0032] Note that a single embodiment may not be limited to detection of cardiac substrate remodeling in only one of atria or ventricles. According to some embodiments, a single embodiment may include a machine learning model that is trained to predict cardiac substrate remodeling in both atria and ventricles. According to such embodiments, the machine learning model may be trained with ECG data labeled as being from patients without a remodeling diagnosis, as well as with ECG data labeled as being from patients with a remodeling diagnosis, where the latter data is further labeled as to whether the diagnosis is of ventricle remodeling or atria remodeling. According to some embodiments, a single embodiment may include multiple machine learning models, e.g., a model that is trained to predict cardiac substrate remodeling in cardiac atria and a model that is trained to predict cardiac substrate remodeling in cardiac ventricles.

[0033] Whether an embodiment can provide cardiac substate remodeling predictions for atria, ventricles, or both, such an embodiment may accept patient input data that includes conventional ECG data. The patient input data may besupplemented with additional patient data, e.g., patient demographic data taken from Electronic Health Records (EHR). For embodiments that include multiple machine learning models, such additional patient data may be included with patient input data for one or more models, and not included with an input of patient data to one or more other models. For example, patient demographic data may be included with ECG data as an input to a model that detects atrium substrate remodeling, but excluded from an input to a model that detects ventricle substrate remodeling, or vice versa.

[0034] By way of non-limiting examples, machine learning models for predicting cardiac substrate remodeling are shown and described herein in reference to Fig. 2 (for cardiac atria) and Fig. 4 (for cardiac ventricles).

[0035] Fig. 2 is a schematic diagram of a machine learning system 200 for assessing cardiac atrial substrate remodeling in a patient using ECG data, according to various embodiments. The system 200 was used for a retrospective analysis of 211 patients with recurrent paroxysmal and persistent (AF duration < 1 year) AF, who underwent redo radiofrequency catheter ablation at a high-volume electrophysiology center (>1000 ablations per year) in Germany. Patient characteristics, including age and comorbidities, were recorded. These patients underwent a previous PVI without substrate modification. Follow-up visits were conducted after three months (the blanking period) and 12 months following the conclusion of the blanking period.

[0036] Training data for the system 200 included labeled ECG data. 12 lead ECGs in Sinus rhythm were acquired 1 -7 days prior to the ablation using Schiller ECG devices (CS 200). The ECG data were then exported as anonymized DICOMs for use by the system 200. Some implementations also used patient demographic data, e.g., height, sex, race, and BMI. According to various implementations, the ECG data (together with associated patient demographic data for some implementations) wereassociated with labels indicative of: existence, extent, and location of LVA, and / or AF classification (e.g., paroxysmal or persistent). The labels for the existence, extent, and location of LVA were determined using electroanatom ical maps (EAMs) to identify representative LVAs, as shown and described herein in reference to Fig. 3.

[0037] Fig. 3 illustrates 3D EAMs acquired from the left atrium before and after excluding the areas of pulmonary veins, mitral valve, and left atrial appendage, generating the low voltage map that serves as a surrogate for atrial remodeling, according to various embodiments. The EAM was performed utilizing the CARTO 3 System while patients were in sinus rhythm. The recording of bipolar voltage amplitudes spanned multiple locations across each atrial segment within the maps. The average voltage signal for each segment, as well as for the entirety of the atrial structure, was then calculated.

[0038] From a total of 209 electroanatom ic maps, 78 high-definition (HD) maps were meticulously chosen for detailed examination to assess the extent of the LVAs. These selected maps had a higher threshold for point collection to enable a more accurate presentation of the substrate. The mapping process was standardized with a fill threshold and color-coding set at level five, ensuring a point density not less than one point per 0.7 cm2. To minimize mapping artifacts, CARTO system’s confidence filters, including cycle length and tissue proximity index, were employed. The point inclusion criterion was set to a five mm radius from the tissue surface to be considered for mapping. Furthermore, the implementation of respiration gating refined the fidelity of the atrial geometry representation. Areas within the atrial walls were then analyzed; regions equal to or greater than 1 cm2exhibiting local signal strength below 0.5 mV were classified as LVAs for subsequent analysis. Following the mapping protocol, the actual surface area of left atria (LA) was ascertained as previously described. Toensure precision, LVAs adjacent to the pulmonary veins (PVs) potentially arising from prior ablations were excluded from the analysis.

[0039] Note that although the system 500 was trained using labeled data where the labels were indicative of cardiac atrial substrate remodeling as detected using EAM to identify LVAs, embodiments are not so limited. Embodiments may be trained with a corpus of labeled training data, where the labels are indicative of atrial substrate remodeling as detected using any, or any combination, of: EAM (e.g., to detect LVAs), LGE MRI, PET, CT, and / or a different modality.

[0040] For data preparation and analysis, the patients used for training the system 200 were stratified according to the extent of LA LVAs. A range of predefined cut-off values for total LVA coverage (5%, 10%, 20%, 30%, and 40%) of LA surface area were used. In assessing the LVA involvement on individual atrial walls, a minimum criterion of 2% of the total left atrial surface area was set.

[0041] To facilitate the use of voltage amplitude data points obtained from the LA, the mean voltage from the aggregate data points across the different atrial walls was computed. Patients for training the system 200 were then categorized into groups based on the average total voltage amplitude, with thresholds established at less than 0.5 mV and up to 1 mV.

[0042] The ECG DICOM files used for training the system 200 were processed using the pydicom package, and ECG signals were extracted. The extracted signals were filtered using a standard bandpass filter and normalized to make sure the amplitude of the signals remain consistent throughout the dataset.

[0043] In addition to the ECG signals, clinical covariates such as age, sex, and BMI were extracted for the patients used to train the system 200. These clinical covariates were scaled using a Robust Scaler. Samples with any missing data wereremoved and only patients were all the data available were used. These processed data variables were used to train the system 200. The system 200 was tested after having been trained using the labeled ECG data both with the clinical covariates (age, sex, and BMI) and without.

[0044] To arrive at the non-limiting example architecture of the system 200 shown and described in reference to Fig. 2, a variety of different architectures were tested. Most of the tested models used architectures based on convolutional neural networks (CNN), long short-term memory (LSTM), and their combinations. After comparing different combinations of these architectures, an architecture that used LSTM and a cross attention module was used to achieve the results presented herein. The architecture used is single layer LSTM with 128 units followed by a standard 1 -layer feedforward network. For the cross-attention model, the raw ECG signals were passed to a one-dimensional CNN and an LSTM. Cross-attention was performed between the extracted features from these two networks. For some implementations tested, scaled clinical covariates were added to the dataset before the feed forward network. Further, all the possible combination of ECG electrodes were tested, and the results presented herein are from the best-performing lead combinations.

[0045] For training, 5-fold stratified cross validation was used to ensure robustness and generalizability of the system 200. All models were trained for 100 epochs with an early stopping strategy. If the model performance did not improve within 10 epochs, the model was saved and the training halted. The learning rate was scheduled using an exponential learning rate scheduler with the initial learning rate being 0.01 , and the exponential decay parameter was 0.9.

[0046] A presentation of the results from the best-performing model according to the system 200 follows. Data are presented as either mean ± standard deviationor, for non-normally distributed variables, as median with interquartile range (IQR). Associations of LA voltage are analyzed using Cox proportional hazard regression across the entire patient cohort. Hazard ratios (HRs) were derived from three distinct Cox regression models: Unadjusted, Adjusted for age and sex, and Adjusted for age, sex, and AF type. All statistical computations were performed using IBM SPSS Statistics Version 26 (IBM Inc., Armonk, NY, USA). A P-value of less than 0.05 was set as the threshold for statistical significance.

[0047] A total of 209 patients were included for analysis. Among these, 69.4% were male, with a mean age of 64 ± 10.7 years. The baseline characteristics of the patients are summarized in Table 1. The majority of the patients had persistent AF (54.8%). The mean BMI was 27.2 ± 4, and 7.6% of the patients had diabetes. A significant portion of the patients had arterial hypertension (aHT) (62.4%), while 11 .4% had obstructive sleep apnea syndrome (OSAS). A similar percentage of patients had experienced a stroke or transient ischemic attack (TIA), accounting for 13.8% of the cohort.

[0048] In Table 1 , continuous variables are presented as mean ± SD or median [interquartile range], depending on their distribution. Categorical variables are presented as n (%). Key: “EF” = Ejection fraction, “BMI” = Body mass index, “TIA” = transient ischemic attack, and “OSAS” = obstructive sleep-apnea syndrome.

[0049] The system 200 displayed an 82.5% accuracy and a 93% sensitivity in predicting the average LA voltage at the 0.7 mV threshold. The system 200 exhibited high accuracy and recall in predicting significant low voltage areas (LVAs), defined as > 5% of the LA surface. Performance metrics for various LVA percentage thresholds are detailed in Table 2.Table 2

[0050] The system 200 successfully predicted the presence of LVAs exceeding2% of the total atrial surface area on each assessed atrial wall (anterior, posterior, and roof of the LA).

[0051] The system 200 demonstrated a high predictive capability using the 12-lead ECG for distinguishing between types of AF, categorized as paroxysmal and persistent.

[0052] Thus, the system 200 not only predicted the presence of LA LVAs with high accuracy, but also discerned the varying degrees and location (posterior wall, anterior wall, or atrial roof) of these LVAs.

[0053] And the system 200 showed that the ECG can anticipate both the types of AF (paroxysmal or persistent).

[0054] The system 200 showed that the optimal leads for predicting LVAs varied depending on the region of the LA, underscoring the nuanced role that the position of the lead plays in diagnosis.

[0055] Detecting LVAs is useful beyond a mere ablation target, extending to a factor in evaluating ablation indications and the potential need for oral anticoagulation medication. The ability to infer details about the atrial substrate from an available, cost-effective tool like the surface ECG could serve as a vital instrument for monitoring LA remodeling in AF patients, thereby influencing the choice of rhythm control strategies or timing for ablation. An ECG indicative of increased LVA percentage may prompt earlier intervention to prevent the progression of atrial myopathy. Together with the ability to prognosticate the lone PVI outcome, the system 200 may be used to enhance AF-patient stratification by enabling personalized ablation strategies and techniques. This may allow physicians to offer tailored therapy and, if necessary,perform additional substrate modification, thereby improving the success rates of ablation procedures.

[0056] Further, the implications of these findings extend to the stratification of risks for cardioembolic events. Given the association between the ECG, the LVAs and cerebrovascular events, ECG data may be analyzed using the techniques disclosed herein to inform the decision as to whether to administer oral anticoagulation.

[0057] Thus, the system 200 demonstrates a successful application of deep learning to 12-lead surface ECG recordings for assessing the LA substrate in AF patients. The system 200 effectively predicts the presence, degree, and location of LVAs with high accuracy, as well as AF classification, offering a novel, non-invasive method for evaluating the LA substrate. Additionally, the system 200 successfully predicts ablation outcomes, underscoring the ECG's potential role in personalized AF management.

[0058] Fig. 4 is a schematic diagram of a machine learning system 400 for assessing cardiac ventricular substrate remodeling in a patient according to various embodiments. The system 400 includes an LSTM layer, which processes the ECG data, a fully connected layer, which processes the output of the LSTM layer and any EHR data (including demographics data as well as ventricular and atrial rates), and an output layer.

[0059] The system 400 was used to assess cardiac ventricular substrate remodeling in a cohort of cardiac sarcoidosis patients. Out of an initial group of n = 391 , 262 patients with ECG were selected, with 240 unpaced and 22 paced. Of those, patients with ECG and LGE MRI images (n=195) were selected and used to train the system 400, with labels indicating whether the LGE MRI images depicted ventricular substrate remodeling in the form of scar tissue visible in the image. The final cohort(n=195) of patients had LGE MRI, ECG and EHR data. Table 3 illustrates the final cohort.Table 3

[0060] The data were cleaned and preprocessed, e.g., by removing records with incomplete data, imputing data, and / or scaling. All combinations of ECG leads were tested. The system 400 performed best using only leads III, V2, and aVF. The best implementation of the system 400 had the following performance metrics: AUC = 0.78, accuracy = 0.775, sensitivity = 0.85, and specificity = 0.7. The system 400 thus predicted the presence of myocardial scar tissue (detectable by LGE-MRI) from ECG signals accurately and with high sensitivity.

[0061] Note that although the system 400 was trained using labeled data where the labels were indicative of myocardial scar tissue as detected using LGE MRI images, embodiments are not so limited. Embodiments may be trained with a corpus of labeled training data, where the labels are indicative of ventricular substrate remodeling as detected using any, or any combination, of: LGE MRI, PET, CT, and / or a different imaging modality.

[0062] Fig. 5 is a flow diagram of a machine learning method 500 of assessing cardiac remodeling in a patient according to various embodiments. The method 500 may be used to assess cardiac substrate remodeling in cardiac atria, cardiacventricles, or both. The method 500 may be implemented using a system such as is shown and described herein in reference to Fig. 2, in reference to Fig. 4, or both.

[0063] At 502, the method 500 includes acquiring patient data that includes patient electrocardiogram (ECG) data. The patient data may be acquired directly from the patient using an ECG machine, or may be acquired from storage in a medical records system, by way of non-limiting examples.

[0064] At 504, the method 500 includes passing the patient data as an input to a trained machine learning model. By way of non-limiting example, the trained machine learning model may include a convolutional neural network, a long short-term memory network, and a cross-attention model. The patient data may be passed to one or more models (e.g., as shown and described herein in reference to Fig. 2 and / or Fig. 4). The trained machine learning model has been trained with a corpus of labeled training data. The corpus of labeled training data includes ECG data, and the labels are indicative of a presence of cardiac remodeling. That is, the labeled training data comes from both normal individuals and individuals for which cardiac substrate remodeling has been detected. The labels may be indicative of remodeling of atria and / or ventricles. According to some embodiments, the patient data may include demographic data for the patient, and the corpus of labeled training data may include demographic data of individuals.

[0065] For atria, the labeled training data may include normal labels and labels representative of cardiac remodeling in atria, and the patient cardiac remodeling indication may be indicative of atrial cardiac remodeling in the patient. According to some embodiments, the labels representative of cardiac remodeling in atria may include labels representative of low voltage areas in atria, and the patient cardiac remodeling indication may be indicative of a low voltage area in a cardiac atrium of thepatient. According to some embodiments, the labels representative of low voltage areas in atria may include labels representative of locations of low voltage areas in atria, and the patient cardiac remodeling indication may be indicative of a location of the low voltage area in the cardiac atrium of the patient. According to some embodiments, the labels representative of cardiac remodeling in atria may include labels representative of atrial fibrillation classification, and the patient cardiac remodeling indication may be indicative of an atrial fibrillation classification for the patient.

[0066] For ventricles, the labeled training data may include normal labels and labels representative of cardiac remodeling in ventricles, and the patient cardiac remodeling indication may be indicative of ventricle cardiac remodeling in the patient.

[0067] At 506, the method 500 includes obtaining a patient cardiac remodeling indication output from the trained machine learning model. The patient cardiac remodeling indication may be indicative of whether or not the patient has cardiac remodeling. The patient cardiac remodeling indication may be obtained through display on a computer monitor, through provision to an electronic health records storage, or by an automated diagnostic medical system, by way of non-limiting examples.

[0068] At 508, the method 500 includes providing an assessment of cardiac remodeling in the patient based on the cardiac remodeling indication. The assessment may be provided through a computer system, or may be provided by a clinician, by way of non-limiting examples.

[0069] Fig. 6 is a flow diagram of method 600 of following up an assessment of atrial cardiac remodeling by the method 500 of Fig. 5, according to variousembodiments. The method may be implemented using a system such as is shown and described herein in reference to Fig. 2, by way of non-limiting example.

[0070] At 602, the method 600 includes obtaining a patient cardiac remodeling indication that is representative of atrium cardiac remodeling. The indication may be obtained from a system 200 as shown and described herein in reference to Fig. 2, using a method such as the method 500 as shown and described in reference to Fig. 5, for example.

[0071] At 604, the method 600 includes determining to treat the patient based on the assessment of cardiac remodeling. The determination may be that the treatment includes cardiac ablation. The patient may be treated 610 after 604, or after proceeding to 606 and / or 608.

[0072] At 606, the method 600 includes determining a location of the cardiac ablation based on the assessment of cardiac remodeling. The patient may be treated 610 after 606, or after proceeding to 608.

[0073] At 608, the method 600 includes determining a type of the cardiac ablation based on the assessment of cardiac remodeling.

[0074] At 610, the method 110 includes providing treatment to the patient as determined by 604, 606, and / or 608.

[0075] Fig. 7 is a flow diagram of method 700 of following up an assessment of ventricle cardiac remodeling by the method 500 of Fig. 5, according to various embodiments. The method may be implemented using a system such as is shown and described herein in reference to Fig. 4, by way of non-limiting example.

[0076] At 702, the method 700 includes obtaining a patient cardiac remodeling indication that is representative of ventricle cardiac remodeling. The indication may be obtained from a system 400 as shown and described herein in reference to Fig. 4,using a method such as the method 500 as shown and described in reference to Fig. 5, for example.

[0077] At 704, the method 700 includes, based on the assessment of cardiac remodeling, imaging a cardiac ventricle of the patient. The imaging may use LGE MRI, echocardiogram, or PET.

[0078] At 706, the method 700 includes predicting heart disfunction based on the imaging.

[0079] At 708, the method 700 includes providing a treatment or prophylaxis to the patient based on the predicting. The treatment or prophylaxis may include one or more of: cardiac ablation, anticoagulant administration, or defibrillation implantation.

[0080] Certain examples can be performed using a computer program or set of programs. The computer programs can exist in a variety of forms both active and inactive. For example, the computer programs can exist as software program (s) comprised of program instructions in source code, object code, executable code or other formats; firmware program(s), or hardware description language (HDL) files. Any of the above can be embodied on a transitory or non-transitory computer readable medium, which include storage devices and signals, in compressed or uncompressed form. Exemplary computer readable storage devices include conventional computer system RAM (random access memory), ROM (read-only memory), EPROM (erasable, programmable ROM), EEPROM (electrically erasable, programmable ROM), flash memory, and magnetic or optical disks or tapes.

[0081] Aspects of the present disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, andcombinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented using computer readable program instructions that are executed by an electronic processor.

[0082] These computer readable program instructions may be provided to a processor of a general-purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the electronic processor of the computer or other programmable data processing apparatus, create means for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks. These computer readable program instructions may also be stored in a computer readable storage medium that can direct a computer, a programmable data processing apparatus, and / or other devices to function in a particular manner, such that the computer readable storage medium having instructions stored therein comprises an article of manufacture including instructions which implement aspects of the function / act specified in the flowchart and / or block diagram block or blocks.

[0083] In embodiments, the computer readable program instructions may be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, statesetting data, configuration data for integrated circuitry, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++, or the like, and procedural programming languages, such as the C programming language or similar programming languages. The computer readable program instructions may execute entirely on a user's computer, partly on the user's computer, as a stand-alone softwarepackage, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.

[0084] As used herein, the terms “A or B” and “A and / or B” are intended to encompass A, B, or {A and B}. Further, the terms “A, B, or C” and “A, B, and / or C” are intended to encompass single items, pairs of items, or all items, that is, all of: A, B, C, {A and B}, {A and C}, {B and C}, and {A and B and C}. The term “or” as used herein means “and / or.”

[0085] As used herein, language such as “at least one of X, Y, and Z,” “at least one of X, Y, or Z,” “at least one or more of X, Y, and Z,” “at least one or more of X, Y, or Z,” “at least one or more of X, Y, and / or Z,” or “at least one of X, Y, and / or Z,” is intended to be inclusive of both a single item (e.g., just X, or just Y, or just Z) and multiple items (e.g., {X and Y}, {X and Z}, {Y and Z}, or {X, Y, and Z}). The phrase “at least one of” and similar phrases are not intended to convey a requirement that each possible item must be present, although each possible item may be present.

[0086] The techniques presented and claimed herein are referenced and applied to material objects and concrete examples of a practical nature that demonstrably improve the present technical field and, as such, are not abstract, intangible or purely theoretical. Further, if any claims appended to the end of this specification contain one or more elements designated as “means for [perform]ing [a function]...” or “step for [performing [a function]...”, it is intended that such elements are to be interpreted under 35 U.S.C. § 112(f). However, for any claims containing elements designated in any other manner, it is intended that such elements are not to be interpreted under 35 U.S.C. § 112(f).

[0087] While the invention has been described with reference to the exemplary examples thereof, those skilled in the art will be able to make various modifications tothe described examples without departing from the true spirit and scope. The terms and descriptions used herein are set forth by way of illustration only and are not meant as limitations. In particular, although the method has been described by examples, the steps of the method can be performed in a different order than illustrated or simultaneously. Those skilled in the art will recognize that these and other variations are possible within the spirit and scope as defined in the following claims and their equivalents.

Claims

What is claimed is:1 . A machine learning method of assessing cardiac remodeling in a patient, the method comprising: acquiring patient data comprising patient electrocardiogram (ECG) data; passing the patient data as an input to a trained machine learning model, wherein the trained machine learning model has been trained with a corpus of training data comprising labels, wherein the corpus of training data comprises ECG data, and wherein the labels are indicative of a presence of cardiac remodeling; obtaining a patient cardiac remodeling indication output from the trained machine learning model, wherein the patient cardiac remodeling indication is indicative of whether the patient has cardiac remodeling; and providing an assessment of cardiac remodeling in the patient based on the cardiac remodeling indication.

2. The method of claim 1 , wherein the labels comprise labels representative of a presence of cardiac remodeling in atria, and wherein the patient cardiac remodeling indication is indicative of atrial cardiac remodeling in the patient.

3. The method of claim 2, wherein the labels comprise labels representative of a presence of low voltage areas in atria, and wherein the patient cardiac remodeling indication is indicative of a low voltage area in a cardiac atrium of the patient.

4. The method of claim 3, wherein the labels comprise labels representative of locations of low voltage areas in atria, and wherein the patient cardiac remodeling indication is indicative of a location of the low voltage area in the cardiac atrium of the patient.

5. The method of claim 2, wherein the labels comprise labels representative of atrial fibrillation classification, and wherein the patient cardiac remodeling indication is indicative of an atrial fibrillation classification for the patient.

6. The method of claim 2, further comprising providing a prognosis of the patient based on the assessment of cardiac remodeling.

7. The method of claim 2, further comprising providing a treatment to the patient based on the assessment of cardiac remodeling, wherein the treatment comprises performing cardiac ablation on the patient.

8. The method of claim 7, further comprising determining a location of the cardiac ablation based on the assessment of cardiac remodeling.

9. The method of claim 8, further comprising determining a type of the cardiac ablation based on the assessment of cardiac remodeling.

10. The method of claim 1 , wherein the labels comprise labels representative of a presence of cardiac remodeling in ventricles, and wherein thepatient cardiac remodeling indication is indicative of ventricle cardiac remodeling in the patient.11 . The method of claim 10, further comprising, based on the assessment of cardiac remodeling: imaging a cardiac ventricle of the patient using at least one of: late gadolinium enhanced magnetic resonance imaging (LGE MRI), echocardiogram, or positron emission tomography (PET); and predicting heart disfunction based on the imaging.

12. The method of claim 11 , further comprising providing a treatment or prophylaxis to the patient based on the predicting, wherein the treatment or prophylaxis comprises at least one of: cardiac ablation, anticoagulant administration, or defibrillation implantation.

13. The method of claim 1 , wherein the patient data further comprises demographic data for the patient, and wherein the corpus of training data further comprises demographic data of individuals.

14. The method of claim 1 , wherein the trained machine learning model comprises a convolutional neural network, a long short-term memory network, and a cross-attention model.

15. The method of claim 1 , further comprising performing a longitudinal study of the patient by repeating the acquiring, the passing, the obtaining, and the providing over a period of time.

16. A machine learning system for assessing cardiac remodeling in a patient, the system comprising a non-transitory computer readable medium comprising instructions, and at least one electronic processor that executes the instructions to perform operations comprising: acquiring patient data comprising patient electrocardiogram (ECG) data; passing the patient data as an input to a trained machine learning model, wherein the trained machine learning model has been trained with a corpus of training data comprising labels, wherein the corpus of training data comprises ECG data, and wherein the labels are indicative of a presence of cardiac remodeling; obtaining a patient cardiac remodeling indication output from the trained machine learning model, wherein the patient cardiac remodeling indication is indicative of whether the patient has cardiac remodeling; and providing an assessment of cardiac remodeling in the patient based on the cardiac remodeling indication.

17. The system of claim 16, wherein the labels comprise labels representative of a presence of cardiac remodeling in atria, and wherein the patient cardiac remodeling indication is indicative of atrial cardiac remodeling in the patient.

18. The system of claim 17, wherein the labels comprise labels representative of a presence of low voltage areas in atria, and wherein the patientcardiac remodeling indication is indicative of a low voltage area in a cardiac atrium of the patient.

19. The system of claim 18, wherein the labels comprise labels representative of locations of low voltage areas in atria, and wherein the patient cardiac remodeling indication is indicative of a location of the low voltage area in the cardiac atrium of the patient.

20. The system of claim 17, wherein the labels comprise labels representative of atrial fibrillation classification, and wherein the patient cardiac remodeling indication is indicative of an atrial fibrillation classification for the patient.21 . The system of claim 16, wherein the labels comprise labels representative of a presence of cardiac remodeling in ventricles, and wherein the patient cardiac remodeling indication is indicative of ventricle cardiac remodeling in the patient.

22. The system of claim 16, wherein the patient data further comprises demographic data for the patient, and wherein the corpus of training data further comprises demographic data of individuals.

23. The system of claim 16, wherein the trained machine learning model comprises a convolutional neural network, a long short-term memory network, and a cross-attention model.

Images