Systems and methods for generating cardiac tissue motion synthetically from surface ECG

Abstract

Methods and systems are provided herein for generating synthetic echocardiography data, such as tissue Doppler imaging data, from signals obtained from electrocardiogram (ECG) leads. Such methods and systems may employ trained generative networks to transform ECG signals into echocardiography signals of a give type, according to user requirements. In some embodiments, assessments of cardiac biomechanical function and determinations of cardiac parameters may be obtained from ECG data, which would traditionally have relied upon real echocardiographic data obtained from ultrasound scans.

Description

Docket No. (183161-37)SYSTEMS AND METHODS FOR GENERATING CARDIAC TISSUE MOTION SYNTHETICALLY FROM SURFACE ECG CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims priority to U.S. provisional patent application number 63 / 764,520 filed February 27, 2025, the entire content of which is incorporated herein by reference.STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH

[0002] N / ABACKGROUND

[0003] Echocardiography is a non-invasive imaging modality that has become a cornerstone in the evaluation of cardiac patients, providing critical assessments of myocardial structure and function, including left ventricular systolic and diastolic performance, tissue motion dynamics, valvular integrity, and chamber geometry. Despite its well-established clinical utility, echocardiography has recognized tradeoffs; in particular, there are several significant limitations that constrain its broader applicability. First, echocardiographic evaluation generally uses specialized ultrasound equipment that, even in some “portable” configurations, is expensive, technically complex, and dependent upon trained sonographers or physicians for proper usage (for operation, proper image acquisition, safety, and results interpretation). Second, echocardiography tends to be limited to an isolated, “snapshot” point-in-time acquisition. Largely due to the cost and complexity of its equipment, echocardiographic studies are typically performed during discrete clinical encounters, and are not suitable for longitudinal, daily life, and / or ‘at-home’ monitoring of cardiac function over extended periods. There is no currently-practiced method by which echocardiographic data can be gathered continuously or repeatedly outside of a clinical setting — such as in a patient's home 14937-9437-8896Docket No. (183161-37) — without the involvement of a trained technician and dedicated imaging equipment. Third, in the context of telemedicine and remote patient care, echocardiography becomes infeasible or even impossible. Even when hand-held or portable echocardiographic devices are available, they still require in-person application by qualified personnel at an appointed time. This constraint is, of course, contrary to the point of remote care.

[0004] Furthermore, the echocardiographic imaging modality has received increasing scrutiny in recent years for the potential of overuse and the associated cost ramifications. Yet, it remains a standard of care for evaluating various types of motion and spatial / image-based cardiac conditions, such as cardiac mechanical function, diastolic performance, and structural abnormalities. No sufficient non-echo imaging alternative has been widely adopted.

[0005] In contrast, surface electrocardiography (ECG) though comparatively low-cost and more portable, has not been adopted for complex motion or spatial-based cardiac evaluation. Rather, surface ECG is generally considered a diagnostic tool that can be readily deployed in ambulatory, remote, and home-based settings, including through wearable consumer devices - but only confined to the assessment of the heart's electrical activity — including cardiac rhythm analysis, conduction abnormality detection, and ischemia screening — rather than the evaluation of mechanical cardiac function or tissue motion. As a result, surface ECG has not been adopted as a substitute for echocardiography in assessing myocardial performance, ventricular mechanics, or structural heart disease, leaving a significant gap in the ability to obtain echocardiographic-equivalent cardiac assessments outside of traditional clinical environments.

[0006] Accordingly, a need exists for a more accessible, remote, patient-supervised, lower cost method for acquiring echocardiographic data in a manner that is clinically useful and overcomes the disadvantages and limits of existing approaches.24937-9437-8896Docket No. (183161-37)SUMMARY

[0007] The following presents a summary of one or more aspects of the present disclosure. This summary is not an extensive overview of all contemplated features of the disclosure and is intended neither to identify key or critical elements of all aspects of the disclosure nor to delineate the scope of any or all aspects of the disclosure. The scope of the disclosure is defined by the claims appended hereto. The descriptions, embodiments, and examples set forth herein are nonlimiting and are provided solely to facilitate understanding of various claims and claim formulations that may be presented.

[0008] In some aspects, the present disclosure provides a system for generating synthetic echocardiographic waveform data from electrocardiogram (ECG) signals. The system comprises at least one ECG electrode configured to be affixed to a surface of a patient and to detect electrical cardiac signals therefrom; a processor; an electronic data connection between the at least one ECG electrode and the processor; an output connection coupled to the processor and configured to convey output data to an output device; and a computer-readable memory coupled to the processor and storing instructions that, when executed by the processor, cause the processor to: receive, via the electronic data connection, the electrical cardiac signals detected by the at least one ECG electrode; preprocess the electrical cardiac signals to generate ECG data comprising at least one segmented cardiac cycle; provide the ECG data as input to a trained generator network, wherein the trained generator network was trained on a training dataset comprising paired ECG data and real tissue Doppler imaging (TDI) waveforms acquired from echocardiographic imaging of a plurality of patients; obtain, from the trained generator network, synthetic echocardiographic waveform data temporally-correlated with actual cardiac tissue movement of the patient over the at least one segmented34937-9437-8896Docket No. (183161-37) cardiac cycle; and provide the temporally-correlated synthetic echocardiographic waveform data in human-interpretable format to the output device via the output connection.

[0009] In some embodiments of the system, the temporally-correlated synthetic echocardiographic waveform data comprises a synthetic tissue Doppler imaging (TDI) waveform temporally-correlated with myocardial tissue velocity of the patient during the at least one segmented cardiac cycle. In some embodiments, the trained generator network was trained as a component of a generative adversarial network (GAN) further comprising: (i) a discriminator network trained to distinguish real TDI waveforms from synthetic TDI waveforms generated by the trained generator network; and (ii) an auxiliary regressor trained to predict peak velocity characteristics of TDI waveforms, wherein the discriminator network and the auxiliary regressor each provided training feedback to the trained generator network during training of the GAN. In some embodiments, the synthetic TDI waveform characterizes a three-peak morphology of the cardiac cycle comprising a systolic velocity peak (s'), an early diastolic velocity peak (e'), and a late diastolic velocity peak (a'). In some embodiments, the instructions further cause the processor to derive synthetic average e' and s' values from the synthetic TDI waveform and generate a mortality risk prediction for the patient based at least in part on the synthetic average e' and s' values, adjusted for one or more of age, sex, presence of atrial fibrillation, and cardiac conduction pattern. In some embodiments, the human-interpretable format comprises a visual depiction of the synthetic echocardiographic waveform data and a graphical indication of a basis for a cardiac state determination corresponding to a clinical diagnostic assessment. In some embodiments, the instructions further cause the processor to extract magnitude and timing measurements from the synthetic echocardiographic waveform data and provide the extracted magnitude and timing measurements as input to a trained machine learning classification model configured to generate a cardiac screening determination based thereon. In some embodiments, the44937-9437-8896Docket No. (183161-37) synthetic echocardiographic waveform data exhibits a cosine similarity of at least 0.75 with corresponding real echocardiographic waveform data acquired from echocardiographic imaging of the patient. In some embodiments, the synthetic echocardiographic waveform data exhibits a linear correlation coefficient (R2) of at least 0.70 with corresponding real echocardiographic waveform data acquired from echocardiographic imaging of the patient.

[0010] In some embodiments of the system, the instructions further cause the processor to derive one or more timing interval measurements from the synthetic TDI waveform, the one or more timing interval measurements comprising at least one of: (a) a time from an onset of an R wave on the ECG data to a peak of a systolic velocity wave on the synthetic TDI waveform; (b) a duration of mechanical contraction determined from the onset of the R wave to a peak displacement in systole, wherein the peak displacement is determined as a time integral of the synthetic TDI waveform; (c) a mechanical relaxation time defined as a difference between an RR interval and the duration of mechanical contraction; (d) a time from the onset of the R wave to an onset of a peak early diastolic velocity wave on the synthetic TDI waveform; or (e) a ratio between a QT interval reflecting electrical systole and the time from the onset of the R wave to the peak displacement reflecting mechanical contraction; wherein the one or more timing interval measurements exhibit a statistically significant linear correlation with corresponding timing interval measurements derived from real TDI waveforms acquired from echocardiographic imaging. In some embodiments, the instructions further cause the processor to integrate the synthetic TDI waveform over a systolic period of the cardiac cycle to derive a synthetic mitral annular plane systolic excursion (MAPSE) value for the patient. In some embodiments, the instructions further cause the processor to estimate a synthetic global longitudinal strain (GLS) value for the patient based at least in part on the synthetic MAPSE value and one or more features derived from the ECG data.54937-9437-8896Docket No. (183161-37)

[0011] In some aspects, the present disclosure provides a method for generating synthetic echocardiographic waveform data. The method comprises acquiring cardiac electrical signals detected at one or more surface electrodes positioned about a cardiac region of a patient and representative of cardiac activity of the patient during a given measurement time period; processing the cardiac electrical signals using a first trained machine-learning model, to generate synthetic echocardiographic waveform data, wherein the synthetic echocardiographic waveform data represents, with at least 80% alignment, non-synthetic echocardiographic waveform data obtained via ultrasound imaging from the patient for the cardiac activity during the given measurement period; and determining, using a second trained machine learning model, a parameter of at least one cardiac biomechanical function from the synthetic echocardiographic waveform data. In some embodiments, the method further comprises generating a digital twin of left ventricular mechanics based on the synthetic echocardiographic waveform data, without use of data obtained via echocardiographic imaging equipment. In some embodiments, the at least one cardiac biomechanical function comprises cardiac tissue motion dynamics, and the second trained machine learning model was trained to determine myocardial tissue velocity characteristics from synthetic echocardiographic waveform training data. In some embodiments, the parameter comprises at least one of tissue Doppler imaging (TDI) velocities, mitral annular plane systolic excursion (MAPSE), or global longitudinal strain (GLS). In some embodiments, the first trained machine-learning model comprises a generator network trained as a component of a generative adversarial network (GAN) further comprising: (i) a discriminator network trained to distinguish real TDI waveforms from synthetic TDI waveforms generated by the trained generator network; and (ii) an auxiliary regressor trained to predict peak velocity characteristics of TDI waveforms, wherein the discriminator network and the auxiliary regressor each provided training feedback to the trained generator network64937-9437-8896Docket No. (183161-37) during training of the GAN. In some embodiments, the at least one cardiac biomechanical function comprises left ventricular mechanics, and the second trained machine learning model was trained to assess at least one of left ventricular systolic dysfunction or left ventricular diastolic dysfunction from the synthetic echocardiographic waveform data. In some embodiments, the left ventricular mechanics further comprise at least one of MAPSE or GLS, and the second trained machine learning model was trained to predict ejection fraction status based at least in part on the MAPSE or GLS derived from the synthetic echocardiographic waveform data. In some embodiments, the at least one cardiac biomechanical function comprises myocardial performance, and the second trained machine learning model was trained to stratify mortality risk for the patient based on parameters derived from the synthetic echocardiographic waveform data and to identify at least one cardiac structural or functional abnormality comprising at least one of diastolic dysfunction, aortic stenosis, valvular heart disease, ventricular hypertrophy, or right ventricular systolic dysfunction from the synthetic echocardiographic waveform data without requiring echocardiographic imaging of the patient. In some embodiments, the method further comprises determining that one or more surface electrodes do not include a desired number or a desired position of electrodes to generate the parameter, based on a predetermined threshold, and synthesizing additional data from missing electrodes using at least one of a transformation equation based on known mathematical relationships between at least two of the one or more surface electrodes that are available, or a vectorcardiographic projection.

[0012] In some aspects, the present disclosure provides a method for assessing cardiac function from electrocardiogram (ECG) data. The method comprises acquiring ECG data from one or more surface electrodes in contact with a patient; processing the ECG data via a foundation model to generate a set of representation features of the ECG data; processing the ECG data through a trained generator network of a generative adversarial network (GAN) to74937-9437-8896Docket No. (183161-37) generate synthetic cardiac velocity waveform data temporally correlated with cardiac tissue movement of the patient; extracting statistical features from the synthetic cardiac velocity waveform data; deriving at least one cardiac mechanical property of the patient from the representation features and the statistical features; and outputting a cardiac function assessment based on the at least one cardiac mechanical property. In some embodiments, deriving the at least one cardiac mechanical property comprises integrating the synthetic cardiac velocity waveform data over a systolic period to derive a synthetic mitral annular plane systolic excursion (MAPSE) value. In some embodiments, the method further comprises estimating a synthetic global longitudinal strain (GLS) value based at least in part on the synthetic MAPSE value and one or more features derived from the ECG data. In some embodiments, deriving the at least one cardiac mechanical property comprises combining predictions from a first machine learning model trained on the statistical features and a second machine learning model trained on the representation features using an ensemble model employing stacking to generate a combined risk score.

[0013] These and other aspects of the disclosure will become more fully understood upon a review of the drawings and the detailed description, which follows. Other aspects, features, and embodiments of the present disclosure will become apparent to those skilled in the art, upon reviewing the following description of specific, example embodiments of the present disclosure in conjunction with the accompanying figures. While features of the present disclosure may be discussed relative to certain embodiments and figures below, all embodiments of the present disclosure can include one or more of the advantageous features discussed herein. In other words, while one or more embodiments may be discussed as having certain advantageous features, one or more of such features may also be used in accordance with the various embodiments of the disclosure discussed herein. Similarly, while example embodiments may be discussed below as devices, systems, or methods84937-9437-8896Docket No. (183161-37) embodiments, it should be understood that such example embodiments can be implemented in various devices, systems, and methods.BRIEF DESCRIPTION OF THE DRAWINGS

[0014] FIG. 1 illustrates a GAN-based model development required pairing ECG and TDI cardiac cycles (left panels), followed by model training with augmented ECG-TDI waveform pairs to enhance data variability, according to some embodiments. Internal and external validation was performed by comparing synthetic TDI waveforms with real TDI waveforms, assessing the similarity and accuracy of synthetic waveforms. Following technical validation, the synthetic TDI waveforms were analyzed for clinical associations (right panel), including correlations with patient demographics and physiological factors such as age, sex, heart rate, and blood pressure. Furthermore, synthetic TDI measurements were used to develop predictive models for screening left ventricular (LV) diastolic dysfunction and ejection fraction (EF) <50%. This process extracts features from synthetic waveforms, such as the magnitude and timings of points (e’, s’, a’). GAN = Generative Adversarial Network, TDI = Tissue Doppler Imaging, ECG = Electrocardiogram, LV = left ventricle, EF = ejection fraction.

[0015] FIG. 2 depicts an overview of the Generative Adversarial Network (GAN) system for runtime and training. GAN produces synthetic TDI waveform from 12-lead ECG. The figure illustrates a process where a patient's 12-lead ECG input (one RR interval) is fed into a feature extractor, which extracts relevant data for generating a TDI waveform. In some embodiments, the extracted data is provided to a GAN or a trained generator network of a GAN and evaluated by a human evaluator for accuracy. This ground truth data is used to train the model, ensuring the generated waveform closely aligns with real-world TDI measurements. GAN = Generative Adversarial Network, TDI = Tissue Doppler Imaging, ECG = Electrocardiogram, RR interval = R-peak to R-peak interval.94937-9437-8896Docket No. (183161-37)

[0016] FIGS. 3A-3D depict real versus synthetic tissue Doppler waveforms. FIG. 3A depicts examples of real and synthetic TDI waveforms obtained from the septal and lateral walls. FIG.3B depicts examples of TDI waveforms averaged across the entire external validation cohort for both real and synthetic measurements. The shaded regions represent the standard deviation, showing the variability of measurements at each time point in the cardiac cycle for the septal and lateral walls. FIG. 3C depicts linear regression analysis demonstrating the strong association between synthetic and real TDI velocity measurements, with an R2value of 0.79 and p-value < 0.0001. FIG. 3D depicts Bland- Altman analysis shows the bias and limits of agreement between synthetic and real TDI measurements, with a mean difference (bias) of -0.23 cm / s and limits of agreement between -3.94 and +3.48 cm / s. TDI = Tissue Doppler Imaging

[0017] FIGS. 4A-4B depict univariate associations with clinical and echocardiography features. Unadjusted odds ratio analyses are presented for FIG. 4A, measured e’ and FIG. 4B, synthetic e’. Error bars represent the 95% confidence interval for the odds ratio. GAN = Generative Adversarial Network, ECG = Electrocardiogram, BMI = body mass index, LVEF = left ventricular ejection fraction, LVDD = left ventricular diastolic dysfunction, VHD = valvular heart disease.

[0018] FIG. 5 depicts the diagnostic value of synthetic TDI waveforms for screening LV dysfunction. The area under receiver operator characteristic curves for a model developed from clinical features and Glasgow ECG interpretation is compared with another model that combines clinical features, Glasgow ECG interpretation, and synthetic TDI measurements for predicting a, LV diastolic and b, systolic dysfunction. TDI = tissue Doppler imaging, BMI = body mass index, LVEF = left ventricular ejection fraction, LV = left ventricular.

[0019] FIGS. 6A-6B depict the survival analysis using TDI-based average e’. Kaplan-Meier survival curves show the cumulative probability of survival for participants stratified by104937-9437-8896Docket No. (183161-37) synthetic TDI-derived average e’ tertiles. FIG. 6A presents data for participants with sinus rhythm, with tertiles defined as follows: Tertile 1 (<6.4 cm / s), Tertile 2 (6.4-9.2 cm / s), and Tertile 3 (>9.2 cm / s). FIG. 6B presents the data for participants with atrial fibrillation, with tertiles defined as Tertile 1 (<4.7 cm / s), Tertile 2 (4.7-6.7 cm / s), and Tertile 3 (>6.7 cm / s). The survival probability is significantly different between the tertiles in both groups, with p-values of <0.0001 for sinus rhythm and 0.002 for atrial fibrillation, indicating that higher average e’ values correlate with better survival outcomes. TDI = tissue Doppler imaging.

[0020] FIGS. 7A-7B depict an example of prediction of high-risk cardiovascular phenotypes and cardiovascular outcomes. FIG. 7A depicts ROC curves in training (AUC = 0.86) and test (AUC = 0.85) cohorts. FIG. 7B presents competing risks analysis of heart failure-related death by predicted risk groups over 15 years (Gray’s test p < .001). AUC = Area under the curve).

[0021] FIG 8. depicts an example ensemble model architecture for diastolic dysfunction prediction. 12-lead ECGs are processed through a foundation model to generate deep learning representations (Pathway 1) and synthetic cardiac velocity waveforms from which statistical features are extracted (Pathway 2). Both pathways feed into tabular ML models that are combined in an ensemble to predict final DD scores. Model developed on n=l,012 and validated on n = 956 patients.

[0022] FIGS. 9A-9C depict an example of incremental value of the ensemble model for diastolic dysfunction detection and risk stratification. FIG. 9A depicts ROC curves comparing an ECG foundation model versus an ensemble model with synthetic cardiac velocity. FIG. 9B depicts odds ratios for risk factors associated with high-risk ECG-detected diastolic dysfunction, including structural markers such as increased LVMi, LAVi, E / e', and reduced e', and clinical factors such as age >65 years, hypertension, diabetes mellitus, and chronic kidney disease (all p<0.0001). FIG. 9C depicts ensemble model diagnostic performance (AUC with 95% CI) across cardiovascular conditions, including aortic stenosis, aortic regurgitation, mitral114937-9437-8896Docket No. (183161-37) regurgitation, reduced LVEF (<50%), left ventricular hypertrophy, tricuspid regurgitation, RV systolic dysfunction, elevated PASP (>45 mmHg), TR velocity (>3.2 m / s), and structural heart disease. AS = aortic stenosis; AR = aortic regurgitation; AUC = area under the curve; CI = confidence interval; CKD = chronic kidney disease; DM = diabetes mellitus; LAVi = left atrial volume index; LVEF = left ventricular ejection fraction; LVMi = left ventricular mass index; MR = mitral regurgitation; NRI = net reclassification improvement; PASP = pulmonary artery systolic pressure; RV = right ventricular; TR = tricuspid regurgitation.

[0023] FIGS. 10A-10C depicts that synthetic LV metrics show directionally consistent associations with clinical characteristics and mirrored trends observed in real TDI measurements. FIG. 10A presents a previously validated generative adversarial network used to generate TDI waveforms from 12-lead ECGs. FIG. 10B presents an example of integration of TDI waveforms to determine the synthetic MAPSE; Machine learning model using echocardiogram and ECG feature inputs to predict LV length and develop synthetic GLS. FIG.10C presents an example of univariate logistic regressions with e’ and clinical characteristics demonstrating similar patters of OR associations between real and synthetic e’ values.

[0024] FIG. 11 depicts a study pipeline outlining the prognostic value of LV mechanics yielding interpretable hazard relationships with synthetic e’ and MAPSE. As used herein, mitral annular plane systolic excursion (MAPSE) refers to the total longitudinal displacement of the mitral annulus during systole, which serves as a surrogate measure of global left ventricular longitudinal systolic function. In some embodiments, synthetic MAPSE is computed by time-integrating the synthetic TDI velocity waveform generated by the trained GAN model (e.g., as generated at step 1108 of method 1100 described in connection with FIG.13) over the systolic period — that is, from the onset of ventricular contraction to the point of peak displacement corresponding to the contraction-relaxation crossover. Because the synthetic TDI velocity waveform represents the instantaneous speed of myocardial longitudinal124937-9437-8896Docket No. (183161-37) motion at each point in the cardiac cycle, integration of this velocity signal over the systolic interval yields a displacement waveform from which MAPSE can be directly measured as the peak systolic displacement value. In some examples, the synthetic MAPSE derived from surface ECG signals via the foregoing method was found to be the strongest multivariate predictor of acute myocardial infarction (AMI) mortality, with each 1-mm increase in synthetic MAPSE associated with markedly lower mortality risk (hazard ratio: 0.19; 95% CI: 0.13-0.26; p < 0.001). Furthermore, the addition of synthetic TDI metrics including MAPSE to conventional ECG analysis improved discrimination for AMI mortality (C- index: 0.70 versus 0.61; p < 0.001), demonstrating significant incremental prognostic value. In a separate validation, a multivariate model using synthetic LV mechanics — including synthetic MAPSE and synthetic global longitudinal strain (GLS) derived from the same ECG-based pipeline — achieved an area under the receiver operating characteristic curve of 0.77 (95% CI: 0.74-0.79) for predicting ejection fraction less than 50% and 0.78 (95% CI: 0.75-0.80) for predicting ejection fraction less than 30%, across cohorts comprising both hospitalized patients with acute myocardial infarction and outpatients with chronic LV dysfunction. These findings demonstrate that synthetic MAPSE and related synthetic LV mechanical parameters derived from surface ECG signals via the methods described herein provide clinically meaningful, interpretable, and scalable biomarkers for both diagnostic assessment of LV systolic dysfunction and prognostic risk stratification for cardiac mortality, in settings where echocardiographic imaging may be unavailable, impractical, or cost-prohibitive.

[0025] FIG. 12 presents a flow chart conceptually illustrating aspects of a method for improving ECG signal information, for example by synthesizing additional ECG lead information.134937-9437-8896Docket No. (183161-37)

[0026] FIG. 13 presents a flowchart illustrating an example of a process for generating clinically relevant echocardiographic information from ECG data, such as parameters that might otherwise be obtained from ultrasound / TDI studies.

[0027] FIG. 14 presents a flowchart illustrating an example process for preparing or training one or more models for generation of echocardiographic information and / or classifications and diagnoses that can be extracted therefrom.

[0028] FIG. 15 depicts a block diagram of an example hardware environment for determining synthetic TDI waveforms and other echocardiogram data from surface ECG leads.DETAILED DESCRIPTION

[0029] The following description will provide a disclosure of various features, approaches, and aspects of example systems and methods that can overcome the limitations described above for acquisition of echocardiographic data, and allow for more practical, patient-supervised, remote, low cost, longitudinal, consistent, less invasive, and accessible approaches. First, a general description will be provided of techniques, algorithms, methods, processes, functionalities and aspects of technologies that may be utilized in systems and methods of the present disclosure. Second, an overview of illustrative system / hardware implementations will be provided along with an overview of a framework for deploying certain processes and algorithms of the present disclosure. Third, a description of the inventors’ experiments and validation studies will be provided.

[0030] Described here are systems and methods directed to improved processing and utilization of ECG signals. Such improvements may be appreciated in a variety of respects: for example, improved acquisition of echocardiographic data (and in some cases including TDI waveform data) from surface ECG readings; improved usage of ECG readings for monitoring, characterization, and / or detection of certain cardiac activity or features that previously could not have been obtained from ECG data only (and in some cases could only 144937-9437-8896Docket No. (183161-37) have been obtained from echocardiographic or ultrasound study data). It should be understood that the systems and methods described below are not limiting of the scope of this disclosure, can be combined in various configurations, and may be adapted to replace, complement, and / or fit with attributes and needs of different clinical or patient care goals and presentations.

[0031] As used herein, the terms "correlated," "temporally-correlated," and "temporally correlated" — when used to describe the relationship between synthetic echocardiographic waveform data and actual cardiac tissue movement or between synthetic and real echocardiographic waveform data — refer to a statistically quantifiable degree of correspondence between the compared waveforms or signals, as measured by one or more accepted statistical similarity metrics. Such metrics include, without limitation: cosine similarity (ranging from -1 to 1, where values closer to 1 indicate greater shape similarity between waveform pairs); linear correlation coefficient (R2), which captures amplitude correlation at each timestep across the cardiac cycle; Bland-Altman analysis, which quantifies the mean bias and limits of agreement between synthetic and real waveform measurements; and any other suitable measure of waveform correspondence such as Pearson correlation coefficient, root mean squared error, or mean absolute error. For example, synthetic echocardiographic waveform data that is "temporally-correlated with actual cardiac tissue movement" refers to synthetic waveform data that exhibits a statistically significant degree of correspondence — as quantified by one or more of the foregoing metrics — with the actual mechanical motion of the patient's cardiac tissue during the same cardiac cycle(s) from which the input ECG data was acquired. Similarly, as used herein, the term "alignment" — when used to describe the degree to which synthetic echocardiographic waveform data represents non-synthetic echocardiographic waveform data (e.g., "at least 80% alignment") — refers to a quantified degree of correspondence between the synthetic and non-synthetic waveforms as154937-9437-8896Docket No. (183161-37) measured by one or more of the foregoing statistical similarity metrics, such as a cosine similarity of at least 0.80, a linear correlation coefficient (R2) of at least 0.80, or an equivalent threshold under another accepted statistical similarity measure, or any combination thereof.Example ECG Signal Expansion Processes

[0032] In some circumstances, it may be the case that having ECG signal data from a comparatively large number, N, of ECG leads would be helpful toward a given patient study or measurement. However, it may also be the case that fewer than N ECG leads are available or practical (for example, because a wearable device does not have n leads, because a patient’s physiology or comfort precludes use of more leads, scarcity of equipment, limited power or edge compute resources, or the patient’s environment is not compatible with more leads such as for surgical patients, athletes, travelers, radiology or radiation therapy patients, etc.). But, in accordance with some aspects of the present disclosure, ECG signals obtained from a comparatively small number of electrodes (1-N, where N is less than the traditional number of electrodes used for a given ECG test / measurement) can be utilized to synthesize multi-lead ECG signals. For example, for a standard 12-lead ECG, which might ordinarily use 10 electrodes, a reduced number of electrodes may be utilized to detect ECG signals (e.g., one channel from one electrode, two channels from two electrodes, etc.) from which a full 12-lead ECG dataset can be derived.

[0033] For example, standard electrocardiography (ECG) for cardiac evaluation traditionally relies on a 12-lead configuration to capture the heart’s electrical activity as recorded from multiple electrode positions on the body surface. The standard 12-lead ECG has been routinely employed for the detection and characterization of cardiac electrical abnormalities, including disturbances of cardiac rhythm such as atrial fibrillation and ventricular tachycardia, conduction abnormalities such as bundle branch blocks and atrioventricular blocks, voltage-164937-9437-8896Docket No. (183161-37) based criteria suggestive of chamber enlargement, ST-segment and T-wave changes associated with myocardial ischemia or injury, and QT interval prolongation. Other forms of ECG utilizing fewer electrodes and fewer leads are also employed in certain clinical and consumer settings — for example, single-lead or 2-lead ECG configurations are commonly found in wearable devices and portable monitors, while 3-lead and 5-lead configurations are frequently used in certain emergency triage environments. However, these reduced-lead configurations have traditionally been limited to basic cardiac rhythm surveillance, heart rate monitoring, and detection of select abnormalities such as atrial fibrillation or gross ST-segment changes suggestive of ischemia, and are not generally regarded as suitable for the sophisticated multiaxis cardiac activity analysis afforded by the full 12-lead configuration (and not as a substitute for echocardiographic data as contemplated herein).

[0034] However, in cases where only a limited number of leads are available, such as with portable devices or wearable monitors, the present disclosure provides computational methods that can be employed to synthetically generate missing leads so as to approximate a larger, multi-lead ECG such as a full 12-lead signal in an improved and reliable way.

[0035] One principle used for deriving multi-lead ECG signals from single- or reduced-lead inputs involves applying transformation equations based on relationships like Einthoven’s triangle. For example, because it is known that the three standard limb leads (I, II, III) of a typical 12-lead ECG are mathematically related, the inventors have found that it is clinically practical to compute any one of these leads via application of algorithms using values for the other two. Similarly, augmented limb leads (aVR, aVL, aVF) can be calculated using detected signals using certain linear transformations.

[0036] Beyond lead synthesis from basic transformations / equations, vectorcardiographic techniques can be used to estimate further ECG signals such as precordial lead signals. Such techniques use vector projections of the heart’s electrical activity in three-dimensional space.174937-9437-8896Docket No. (183161-37)

[0037] More advanced techniques are also contemplated which may utilize machine learning algorithms to reconstruct missing leads. This can be particularly useful when reconstructing missing leads from ECG signals that are merely single-lead or have multiple leads but not enough related limb / torso leads to transform signals into missing leads. These models can be developed by training machine learning algorithms on large datasets containing both complete multi-lead ECG signals and their corresponding subsets, allowing the algorithms to learn complex relationships between different leads / lead values. Deep learning techniques, such as generative adversarial networks (GANs) or convolutional neural networks (CNNs), can synthesize missing ECG signals based on the time-series patterns of lead signals that are available, effectively enhancing diagnostic capabilities in resource-limited environments.

[0038] Missing ECG signal reconstruction can be further improved using various processing techniques such as adaptive filtering and signal processing techniques to compensate for noise and artifacts, before applying lead transformation algorithms. Similarly, data reduction and / or feature extraction methods, such as principal component analysis (PCA), can be used to further refine available / recorded signals by identifying key patterns in the existing data, which can be used by ML models to extrapolate them into additional or missing leads.

[0039] Referring now to FIG. 12, an example method 1000 is illustrated for improving diagnostic power of a limited-lead ECG data acquisition. In one aspect, the method 1000 can be utilized to synthesize signals that would have been generated by ECG electrodes that were not actually physically present, were not actually being utilized, or otherwise were not actually generating usable, valid, or sufficient signals. In another aspect, the method 1000 can be thought of as being usable to correct, de-noise, or improve accuracy of leads that are present by interpolating an expected signal based on other leads. In embodiments where method 1000 utilizes trained machine learning models (such as GANs or CNNs) for synthesizing missing ECG lead signals, such models may be developed by training on large datasets of complete184937-9437-8896Docket No. (183161-37) multi-lead ECG recordings. For example, a training dataset comprising a plurality of complete 12-lead ECG recordings may be assembled, and during training, one or more leads may be systematically withheld or masked from each recording to serve as ground truth targets, while the remaining leads are provided as input to the model. The model is then trained to reconstruct the withheld lead signals from the available lead signals, thereby learning the complex spatial and temporal relationships between different ECG leads. Training may employ supervised learning objectives such as reconstruction loss (e.g., LI or L2 loss) between the predicted and actual withheld lead signals, and may further incorporate adversarial training (e.g., via a discriminator network that distinguishes real from synthesized lead signals) to improve the fidelity and realism of the reconstructed waveforms. Data augmentation techniques — such as random lead dropout, noise injection, phase shifting, and amplitude scaling — may be applied during training to improve model robustness across varying lead configurations, signal quality levels, and patient populations. The trained model may be validated on held-out datasets, including recordings from independent clinical sites, to confirm generalizability prior to deployment in method 1000.

[0040] At block 1002, the method 1000 determines the availability of ECG signals. For example, the method 1000 may identify which ECG electrodes are physically present and operational, assess the quality and validity of signals being generated by each available electrode, and determine the number and type of ECG leads for which usable signal data currently exists. In some embodiments, this determination may include identifying whether the available signals originate from limb leads, precordial leads, augmented leads, or any subset thereof, as well as evaluating signal quality metrics such as signal-to-noise ratio, baseline wander, and artifact contamination. In further embodiments, the method 1000 may receive information regarding the type of ECG acquisition device (e.g., a wearable device, a portable monitor, a clinical-grade ECG system) and the electrode configuration associated therewith.194937-9437-8896Docket No. (183161-37)

[0041] In some embodiments, at block 1002, the method 1000 may further include confirming the identity and placement of available ECG electrodes. For example, where the ECG data acquisition is patient-supervised — such as in a home-based, remote, or telemedicine setting in which the patient, a caregiver, a family member, or another non-clinical individual is responsible for electrode placement — the method 1000 may present a guided graphical depiction of the human body on a display (e.g., users’ mobile device, computer, or other output device 1308) illustrating recommended electrode placement locations, and prompt the user to confirm that each electrode has been placed at the indicated position. In further embodiments, the method 1000 may utilize augmented reality (AR) overlays or the camera of a mobile device (e.g., a smartphone or tablet) to visually guide the user through electrode placement and to verify, via image recognition or fiducial marker detection, that electrodes are positioned at the correct anatomical landmarks. Upon confirmation of electrode placement, the method 1000 may map each signal channel to its corresponding lead location, thereby establishing which signal channels correspond to which anatomical electrode positions. Alternatively, in embodiments where the ECG acquisition device has a known, fixed electrode configuration — such as a wearable device designed to be worn at a predetermined location (e.g., a wrist, an upper arm, or a chest-worn patch) — the method 1000 may rely on pre-programmed assumptions that associate specific signal channels with specific lead locations based on the device's intended wearing position, without requiring active user confirmation of electrode placement. In such cases, the device firmware or associated software may be configured with a default mapping of signal channels to anatomical lead positions corresponding to the device's recommended use.

[0042] At block 1004, the method 1000 determines a desired number N of ECG signals or leads and compares the desired number N to the number and type of ECG signals determined to be available at block 1002. Based on this comparison, the method 1000 identifies which lead204937-9437-8896Docket No. (183161-37) signals, if any, are missing, insufficient, or otherwise need to be synthesized. For example, if a full 12-lead ECG dataset is desired but only a single-lead or reduced-lead subset is available, the method 1000 determines the specific leads for which synthetic signals should be generated. In some embodiments, the desired number N may be directly specified by a user through a user interface. In other embodiments, the desired number N may be determined automatically based on the type of analysis or study for which the ECG signals are intended. For example, if a comprehensive ECG evaluation of cardiac electrical activity is requested, the method 1000 may determine thatN should correspond to a full 12-lead configuration at standard 12-lead electrode locations. Conversely, if only a particular type of evaluation is needed that can be accomplished with fewer than 12 leads, the method 1000 may set N accordingly so as to avoid unnecessary synthesis. In embodiments where a three-dimensional cardiac model, tissue motion information, or other echocardiographic-equivalent data is desired — such as the signal-to-signal transformation processes described herein — the method 1000 may determine which N lead locations and configurations will optimally provide the data required for such downstream analyses. In some embodiments, based on the foregoing determination, the method 1000 may further prompt the user to place additional physical electrodes at specified locations and / or to reposition existing electrodes to alternative locations so as to optimize the quality and completeness of the synthesized lead signals.

[0043] At block 1006, the method 1000 enters an iterative loop in which, for each lead signal identified at block 1004 as requiring synthesis, the method 1000 determines an appropriate synthesis technique based on the lead signals available from actual, physically present electrodes. For a given lead to be synthesized, the method 1000 may select from among one or more of the following synthesis approaches:

[0044] (a) Transformation equations based on known mathematical relationships between leads. For example, because the three standard limb leads (I, n, and III) of a typical 12-lead214937-9437-8896Docket No. (183161-37) ECG are mathematically related through Einthoven's triangle, any one of these leads can be computed by applying algorithms using values from the other two. Similarly, augmented limb leads (aVR, aVL, and aVF) can be derived from the standard limb leads using known linear transformations. Accordingly, where at least two of the three standard limb leads are available from physical electrodes, the method 1000 may apply these transformation equations to compute the missing limb lead(s) and augmented limb lead(s).

[0045] (b) Vectorcardiographic projection techniques. Where available lead signals permit estimation of the heart's electrical activity as a vector in three-dimensional space, the method 1000 may employ vectorcardiographic projection techniques to estimate additional lead signals, such as precordial lead signals, by projecting the cardiac electrical vector onto the axes corresponding to the desired lead positions. This approach can be particularly useful for synthesizing precordial leads (V1-V6) from available limb lead data, as the precordial leads represent projections of the cardiac vector onto the transverse plane at specific anatomical locations across the chest.

[0046] (c) Trained machine learning models. The method 1000 may utilize trained machine learning models, such as generative adversarial networks (GANs) or convolutional neural networks (CNNs), to reconstruct missing lead signals based on time-series patterns learned from large datasets of complete multi-lead ECG recordings. These models are trained on datasets containing both complete multi-lead ECG signals and their corresponding subsets, enabling the models to learn complex, non-linear relationships between different leads that may not be captured by deterministic transformation equations or vectorcardiographic projections. This approach can be particularly useful when reconstructing missing leads from ECG signals that are single-lead or that have multiple leads but lack sufficient related limb or torso leads to permit deterministic transformation into missing leads. In some embodiments, prior to applying the machine learning model, the method 1000 may further employ adaptive filtering and signal224937-9437-8896Docket No. (183161-37) processing techniques to compensate for noise and artifacts in the available lead signals, and may apply data reduction and feature extraction methods, such as principal component analysis (PCA), to identify key patterns in the available signals that can be used by the machine learning model to extrapolate them into the missing leads.

[0047] In some embodiments, the selection of synthesis technique for a given lead may depend on which leads are available from physical electrodes, the type and quality of the available signals, and the identity of the lead to be synthesized. For example, where deterministic mathematical relationships exist between available leads and the lead to be synthesized (such as among limb leads I, n, and III), the method 1000 may preferentially apply transformation equations; where such deterministic relationships are not available but sufficient vectorcardiographic data can be derived, the method 1000 may employ vectorcardiographic projection; and where neither deterministic transformation nor vectorcardiographic projection is feasible or sufficient, the method 1000 may utilize trained machine learning models. In further embodiments, the method 1000 may combine multiple synthesis techniques — for example, using transformation equations to derive certain limb leads and then using the derived leads together with the originally available leads as inputs to a machine learning model for synthesizing precordial leads.

[0048] At block 1008, method 1000 then applies the selected technique to generate the given ECG lead signal to be synthesized. This may be done continuously throughout an analysis, at periodic points in time, for given periodic or sequential time windows, etc. In some embodiments, method 1000 may simply store acquired signals during a given time period, and then offload the stored signal data for subsequent calculation and synthesis as described herein. In other embodiments, the synthesis technique(s) may be applied in real time, continuously, automatically, or at predetermined intervals on device or remotely.234937-9437-8896Docket No. (183161-37)

[0049] At block 1010, method 1000 determines whether additional lead information should be synthesized. This may include validation of already-synthesized lead signals and / or determination that more synthesized leads still need to be generated. If additional synthesis is needed, method 1000 may return to block 1006. If not, method 1000 may proceed to block 1012.

[0050] At block 1012, upon completion of the iterative loop of block 1006-1010 for all leads requiring synthesis, the method 1000 outputs a set of N ECG signals corresponding to the desired lead configuration. The output set of N ECG signals may comprise a combination of signals obtained directly from physically present electrodes and synthetically generated signals for leads that were not physically present, were not operational, or did not produce signals of sufficient quality. In some embodiments, the N output signals may not overlap with the originally available physical lead signals — for example, where the method 1000 determines that a synthetically reconstructed version of an available lead (e.g., generated by interpolation from other leads) is of higher quality or consistency than the originally detected signal, the synthetic version may be substituted. In other embodiments, the output set of N ECG signals may be provided as input to downstream processes, such as the signal-to-signal transformation processes described herein for generating synthetic echocardiography signals, clinical decision support models, or other diagnostic or monitoring applications.Example Signal-to-Signal Transformation Processes

[0051] In some circumstances, it may be desirable to obtain information about cardiac tissue motion, myocardial velocity, or other mechanical parameters of the heart — information that has traditionally required echocardiographic imaging, such as Tissue Doppler Imaging (TDI), to acquire. However, echocardiographic assessment may be unavailable or impractical in many clinical and non-clinical settings, including remote or home-based patient monitoring,244937-9437-8896Docket No. (183161-37) telemedicine encounters, ambulatory or wearable device environments, resource-limited facilities, and situations requiring longitudinal or continuous cardiac assessment without repeated in-person imaging studies. In such circumstances, surface ECG signals may be the only cardiac data readily available, yet surface ECG has not traditionally been capable of providing information about cardiac tissue motion or mechanical function. Accordingly, in a second type of transformation, the present disclosure provides methods for using ECG signal information to generate synthetic echocardiography signals, such as TDI waveforms, that represent the speed of myocardial motion, blood motion, and other mechanical parameters such as deformation (strain). The generated parameters can be seeded into a still or dynamic image model of the cardiac chamber (generic mathematical or numeric models or individualized Al models) to give it motion attributions that lead to developing a digital twin of a beating heart. Conversely, the ECG-generated speed of myocardial motion can be used to develop high-speed cardiac ultrasound images by allowing physiological synthetic data to be interpolated in between real frames of images to create high frame rate cardiac ultrasound imaging. A TDI waveform in traditional echocardiography refers to a waveform representing velocity or motion of myocardial tissue, that is derived from Doppler-interpreted ultrasound data. In other words, TDI waveforms show changes in velocity (e.g., movement, acceleration, etc.) of cardiac tissue, which are measured by Doppler shifts in echoes detected by ultrasound transducers. (Note, TDI waveforms differ from other types of echocardiography signals, which typically measure blood flow rather than tissue movement). TDI waveforms can be used to assess myocardial contraction and relaxation dynamics, helping in the evaluation of various cardiac structural and functional details, including diastolic function, ventricular synchrony, and myocardial performance. TDI waveforms generated through the methods presented herein may also comprise any of the several variants or subsets of TDI waveform information, such as: pulsed wave TDI, color TDI, and continuous wave TDI. Furthermore, other types of tissue-254937-9437-8896Docket No. (183161-37) motion waveforms that can be obtained from echocardiography can also be synthesized by the methods herein, such as precise-location specific echocardiography like M-Mode. By way of illustration, an example process for developing a model that will synthesize TDI waveform information from ECG signals may include the steps as outlined in FIG. 13.

[0052] FIG. 13 is a flowchart illustrating an example of a method 1100 for generating echocardiographic data from surface ECG signals obtained from a given patient wearing the associated ECG electrodes in a given placement. As described below, a particular implementation of method 1100 can omit some or all illustrated features or steps, may be implemented in some embodiments in a different order, and may not require some illustrated features to implement all embodiments. In some examples, an apparatus (e.g., processor 1304 with memory 1306 of a computing device 1302 as described in connection with FIG. 15) can be used to perform method 1100. However, it should be appreciated that any suitable apparatus or means for carrying out the operations or features described below may perform method 1100.

[0053] In some embodiments, method 1100 can be used to generate synthetic TDI waveforms from ECG signals and to derive clinically relevant parameters therefrom, such as systolic velocity (s'), early diastolic velocity (e'), late diastolic velocity (a'), mitral annular plane systolic excursion (MAPSE), global longitudinal strain (GLS), or other indices of ventricular mechanics. In some embodiments, synthetic GLS is computed by dividing the synthetic MAPSE value — derived by time-integration of the synthetic TDI velocity waveform over the systolic period — by a predicted LV longitudinal length generated by a separately trained machine learning model, with the resulting ratio expressed as a percentage representing longitudinal myocardial shortening during systole. In other embodiments, method 1100 can be used to determine diagnostic, prognostic, or risk-stratification information for a patient, including assessment of systolic dysfunction, diastolic dysfunction, structural heart disease, or264937-9437-8896Docket No. (183161-37) mortality risk, based on ECG data. In some embodiments, the output of method 1100 can be used to guide a medical provider’s choice of care or to prescribe a particular treatment. In other embodiments, the output of method 1100 can provide information to an insurance company regarding insurance coverage decisions for a patient. In further embodiments, the output of method 1100 can be used to generate or update a digital twin model of a patient's heart, to interpolate high-frame-rate cardiac ultrasound images, or to provide longitudinal monitoring of cardiac mechanical function over time.

[0054] Accordingly, in some embodiments, an additional step or phase may be performed in parallel with or after performance of methods such as shown in FIG. 13. In these embodiments, the synthetic TDI waveform data generated by method 1100 — including synthetic tissue velocity waveforms and parameters derived therefrom such as s', e', a', MAPSE, and GLS — is utilized to predict, classify, or determine one or more cardiac conditions or risk states that previously required actual echocardiographic ultrasound imaging to assess. For example, as described in further detail herein, the synthetic TDI waveform data may be utilized to predict diastolic dysfunction risk phenotypes (as illustrated in connection with FIG. 8), to predict left ventricular systolic dysfunction (as illustrated in connection with FIG. 10), and to predict cardiac mortality risk (as illustrated in connection with FIG. 11). Previously, such predictions, classifications, and diagnostic determinations could only be made from tissue Doppler imaging waveforms and other echocardiographic measurements obtained via actual ultrasound scans performed by trained sonographers using specialized imaging equipment. By generating synthetic echocardiographic motion data from surface ECG signals via method 1100, these diagnostic and prognostic capabilities are extended to settings in which echocardiographic imaging is unavailable, impractical, or cost-prohibitive, including remote, ambulatory, homebased, and resource-limited environments.274937-9437-8896Docket No. (183161-37)

[0055] At block 1102, method 1100 determines a desired echocardiographic output or analysis type. For example, at block 1102, a user or an automated system may specify the type of echocardiographic information to be generated from the ECG signals. In some embodiments, the desired output may comprise a synthetic TDI waveform representing myocardial tissue velocity over a cardiac cycle, from which parameters such as e', s', and a' can be extracted. In other embodiments, the desired output may comprise a diagnostic classification or prediction, such as a determination of whether the patient exhibits LV diastolic dysfunction, LV systolic dysfunction (e.g., ejection fraction less than 50%), or other cardiac abnormalities. In further embodiments, the desired output may comprise a prognostic risk assessment, such as a mortality risk stratification based on synthetic TDI-derived parameters. In still further embodiments, the desired output may comprise a three-dimensional cardiac model, a digital twin representation, or high-frame-rate cardiac imaging data. The determination at block 1102 may be received via a user interface, may be pre-configured based on the deployment environment (e.g., a wearable device configured for continuous diastolic function monitoring), or may be determined automatically based on the clinical context or available data.

[0056] At block 1104, method 1100 assesses whether sufficient ECG lead information is available to support the desired echocardiographic output determined at block 1102. For example, certain echocardiographic outputs may require a full 12-lead ECG dataset as input to the trained model, while other outputs may be achievable with fewer leads. At block 1104, method 1100 may evaluate the number, type, and qualify of ECG lead signals currently available — for example, by identifying which electrodes are physically present and operational, assessing signal qualify metrics such as signal-to-noise ratio, and determining whether the available lead configuration is sufficient for the desired output. If the available ECG lead information is determined to be insufficient — for example, because fewer than the required number N of leads are available from physical electrodes — method 1100 may invoke284937-9437-8896Docket No. (183161-37) method 1000 (as described in connection with FIG. 12) to synthesize additional ECG lead signals from the available signals, thereby augmenting the ECG dataset to the required lead configuration. Upon completion of block 1104, method 1100 proceeds with a set of ECG lead signals (whether obtained directly from physical electrodes, synthesized via method 1000, or a combination thereof) that is sufficient for the desired echocardiographic output.

[0057] At block 1106, method 1100 preprocesses the ECG signals for input to the trained model. In some embodiments, preprocessing may include dividing the ECG signals into individual cardiac cycles based on RR intervals, resizing each cycle to a standardized length (e.g., 512 timesteps), and extracting numerical features from the ECG signals. Such numerical features may include, for example, basic intervals (RR, PR, QT), wave-specific attributes (axis, amplitude, duration), heart rate variability measures, morphological characteristics, dispersion measures, and corrected QT intervals across multiple leads. In some embodiments, the extracted features may be normalized (e.g., by subtracting the mean and scaling to unit variance based on parameters established during model training). In further embodiments, where multiple cardiac cycles are available, method 1100 may select one or more representative cycles for processing, or may process multiple cycles and average the resulting outputs to improve robustness and reduce the effect of inter-beat variability. In some embodiments, preprocessing may further include adaptive filtering, signal smoothing, or artifact removal to improve signal quality prior to model input.

[0058] At block 1108, method 1100 applies a trained generative model to the preprocessed ECG signals to generate a synthetic echocardiographic output. In some embodiments, the trained generative model comprises a GAN (such as a conditional GAN comprising a generator network) that has been previously trained on paired ECG and TDI waveform data, as described herein. At block 1108, the preprocessed ECG waveform data and extracted features are provided as input to produce a synthetic TDI waveform. In some embodiments, the ECG data294937-9437-8896Docket No. (183161-37) is provided to a GAN or a trained generator network of a GAN, which produces the synthetic TDI waveform (TDIGAN) representing the predicted myocardial tissue velocity over the cardiac cycle. In some embodiments, separate trained models may be applied for different measurement locations — for example, a first model trained to generate synthetic TDI waveforms corresponding to the septal wall and a second model trained to generate synthetic TDI waveforms corresponding to the lateral wall. In such embodiments, the outputs of the separate models may be averaged or otherwise combined in accordance with clinical guidelines (e.g., averaging septal and lateral e' values). In other embodiments, the trained model may generate other types of synthetic echocardiographic data, such as M-Mode waveforms, strain curves, or velocity data at other myocardial locations.

[0059] At block 1110, method 1100 extracts clinically relevant parameters from the synthetic echocardiographic output generated at block 1108. In some embodiments, where the output of block 1108 comprises a synthetic TDI waveform, block 1110 may include identifying and measuring the magnitudes and timings of characteristic waveform features, including the peak systolic velocity (s'), the peak early diastolic velocity (e'), and the peak late diastolic velocity (a'). In further embodiments, block 1110 may include deriving interval-based measurements from the synthetic TDI waveform, such as the time from the onset of the R wave on the ECG to the peak of the systolic velocity wave, the duration of mechanical contraction (determined as a time integral of the TDI velocity waveform), the mechanical relaxation time, and the ratio of the QT interval to the time to peak displacement. In some embodiments, block 1110 may further include computing derived metrics such as MAPSE (e.g., by integrating the synthetic TDI velocity waveform over the systolic period to obtain a displacement waveform and measuring the peak systolic displacement value therefrom) and / or synthetic GLS. In some embodiments, synthetic GLS is computed by first determining MAPSE from the synthetic TDI velocity waveform as described herein, and then dividing the MAPSE value by a predicted left304937-9437-8896Docket No. (183161-37) ventricular longitudinal length to obtain a strain value expressed as a percentage. The predicted LV longitudinal length may be generated by a trained machine learning model that receives as input one or more of: features extracted from the synthetic TDI waveform, features extracted from the ECG signals, and patient demographic or clinical features (e.g., age, sex, body surface area). In some embodiments, the machine learning model for LV length prediction is trained on a dataset comprising paired ECG recordings and echocardiographic measurements in which LV longitudinal length is measured from standard echocardiographic imaging views (e.g., apical four-chamber or apical two-chamber views), such that the trained model learns to estimate LV length from ECG-derived inputs without requiring echocardiographic imaging at runtime. In embodiments where multiple cardiac cycles have been processed, block 1110 may include averaging the extracted parameters across cycles.

[0060] As used herein, synthetic global longitudinal strain (GLS) refers to a measure of left ventricular myocardial deformation derived entirely from surface ECG signals via the synthetic TDI pipeline described herein, without requiring echocardiographic imaging at the time of assessment. In traditional echocardiography, GLS is obtained by speckle tracking analysis of two-dimensional echocardiographic images and represents the percentage of longitudinal shortening of the left ventricular myocardium during systole, serving as a sensitive marker of subclinical myocardial dysfunction that may be abnormal even when ejection fraction is preserved. In the synthetic GLS computation described herein, the longitudinal shortening component is captured by the synthetic MAPSE value — which represents the peak systolic displacement of the mitral annulus derived by time-integration of the synthetic TDI velocity waveform — and the reference length component is provided by a predicted LV longitudinal length generated by a separately trained machine learning model. Synthetic GLS is then computed as the ratio of MAPSE to the predicted LV length, expressed as a negative percentage consistent with the convention used in echocardiographic strain imaging (wherein negative314937-9437-8896Docket No. (183161-37) values denote myocardial shortening). In some examples, the synthetic GLS derived from surface ECG signals via the foregoing method, when used together with other synthetic LV mechanical parameters including synthetic MAPSE and synthetic TDI velocities in a multivariate model, achieved an area under the receiver operating characteristic curve of 0.77 (95% CI: 0.74-0.79) for predicting ejection fraction less than 50% and 0.78 (95% CI: 0.75-0.80) for predicting ejection fraction less than 30%, across cohorts comprising 1,410 patients including both hospitalized patients with acute myocardial infarction and outpatients with chronic LV dysfunction. These findings demonstrate that synthetic GLS derived from surface ECG signals provides a clinically meaningful and scalable biomarker for assessing LV systolic function in settings where echocardiographic speckle tracking analysis is unavailable, impractical, or cost-prohibitive.

[0061] At block 1112, method 1100 may optionally apply one or more trained classifier or predictive models to the parameters extracted at block 1110 to generate diagnostic, prognostic, or risk-stratification determinations. For example, in some embodiments, a trained classifier model (such as a naive Bayes model, logistic regression model, support vector machine, random forest, or other machine learning model) may receive as input the synthetic TDI-derived parameters (e.g., e', s', a', and interval-based features), optionally in combination with clinical features (e.g., age, blood pressure, comorbidities) and / or conventional ECG interpretation data, and may output a prediction regarding the presence or absence of LV diastolic dysfunction, LV systolic dysfunction (e.g., EF less than 50%), or other cardiac conditions. In other embodiments, the synthetic TDI-derived parameters may be used as inputs to a prognostic model that stratifies the patient's mortality risk or predicts cardiovascular outcomes based on the extracted parameters. In further embodiments, block 1112 may be omitted where the desired output determined at block 1102 comprises the synthetic TDI324937-9437-8896Docket No. (183161-37) waveform or extracted parameters themselves, rather than a diagnostic or prognostic determination.

[0062] At block 1114, method 1100 outputs the desired echocardiographic data to a user or to a downstream system. In some embodiments, the output may comprise the synthetic TDI waveform generated at block 1108, displayed as a waveform plot on an output device 116 (e.g., a display of a computing device 110, a mobile device, or a clinical workstation) in a format consistent with conventional TDI waveform presentation. In other embodiments, the output may comprise the clinically relevant parameters extracted at block 1110, presented in a tabular or graphical format. In further embodiments, the output may comprise a diagnostic or prognostic determination generated at block 1112, such as a classification of the patient's diastolic or systolic function status, a risk score, or a recommendation for further echocardiographic evaluation. In some embodiments, the output may be transmitted over a communication network 130 to a remote clinician, an electronic medical record system, a telemedicine platform, or a clinical decision support system. In other embodiments, the output may be stored locally on the computing device 110 for longitudinal tracking and comparison with prior or subsequent assessments. In further embodiments, the output may be provided as input to a three-dimensional cardiac modeling application, a digital twin framework, or a high-frame-rate cardiac imaging reconstruction system. In some embodiments, the human-interpretable format in which the synthetic echocardiographic waveform data is provided to the output device may comprise a visual depiction of the synthetic echocardiographic waveform data — such as a waveform plot, a graphical overlay, or a numerical summary — together with a graphical indication of a basis for a cardiac state determination corresponding to a clinical diagnostic assessment, such as a highlighted region of the waveform indicating an abnormal peak velocity, a color-coded risk indicator, or an annotated comparison of extracted parameters against clinical reference thresholds. In some embodiments, the output may further comprise a334937-9437-8896Docket No. (183161-37) cardiac screening determination generated by a trained machine learning classification model that receives as input the magnitude and timing measurements extracted from the synthetic echocardiographic waveform data at block 1110, wherein the cardiac screening determination indicates whether the patient's synthetic echocardiographic parameters meet or exceed one or more clinical thresholds associated with a cardiac condition such as LV diastolic dysfunction, LV systolic dysfunction, or structural heart disease, thereby facilitating a determination of whether referral for further echocardiographic evaluation is warranted. In some embodiments, method 1100 may be repeated at periodic intervals, continuously, or upon detection of a change in the patient's ECG signals, thereby enabling longitudinal monitoring of cardiac mechanical function without repeated echocardiographic imaging.

[0063] Referring now to FIG. 14, a flowchart is illustrated depicting aspects of a method 1200 for development, preparation, and / or validation of one or more models for use in a method such as method 1100 of FIG. 13 and / or the example method of FIG. 8. Method 1200 may be performed offline — that is, prior to deployment of the trained model(s) for runtime inference on patient ECG data — and may be repeated or refined as additional training data becomes available or as model objectives change. As described below, method 1200 includes blocks for developing and validating a GAN model for transforming ECG signals into synthetic echocardiography waveform data, and optionally includes additional blocks for developing a foundation ECG model and / or downstream classifier or predictive models. In some examples, an apparatus (e.g., processor 1304 with memory 1306 of a computing device 1302 as described in connection with FIG. 15) can be used to perform method 1200. However, it should be appreciated that any suitable apparatus or means for carrying out the operations or features described below may perform method 1200.

[0064] At block 1202, method 1200 determines one or more objectives for which the model(s) being developed will be used (e.g., determining inputs / outputs, use cases, user preferences,344937-9437-8896Docket No. (183161-37) etc.) and selects one or more appropriate model architectures. For example, at block 1202, a developer or automated system may determine a desired signal-to-signal transformation — such as a transformation from 12-lead, N-lead, or fewer-than-N lead ECG signals to TDI waveforms representing myocardial tissue velocity — and may select a GAN architecture suitable for the specified transformation. In some embodiments, the selected architecture may comprise a conditional GAN comprising a generator network (G) that produces synthetic TDI waveforms from ECG signal channels as inputs, a discriminator network (D) that distinguishes real TDI waveforms from synthetic TDI waveforms, and optionally an auxiliary regressor (A) that predicts key waveform characteristics to provide additional training feedback. In some embodiments, separate model instances may be configured for different measurement locations — for example, a first model instance for generating synthetic TDI waveforms corresponding to the septal wall and a second model instance for generating synthetic TDI waveforms corresponding to the lateral wall. At block 1202, initial hyperparameters may also be set, including learning rates, loss function weights, regularization parameters, training epoch count, and optimizer configurations for each model component. The selection of the pix2pix conditional GAN architecture and the specific hyperparameter configuration described herein was informed by systematic evaluation of alternative approaches, as described below.

[0065] In some examples, the pix2pix conditional GAN architecture was selected because the underlying task — transforming a 12-lead ECG waveform into a corresponding TDI velocity waveform — constitutes a paired signal-to-signal translation problem, for which conditional GANs designed for paired input-to-output mapping are particularly well suited. Alternative generative architectures were considered, including CycleGAN-based approaches designed for unpaired image-to-image translation; however, because the prospective study design ensured that ECG and echocardiographic recordings were obtained contemporaneously from the same patients (preferably within three hours of each other), the availability of well-paired training354937-9437-8896Docket No. (183161-37) data made the pix2pix conditional framework the most appropriate choice. The three-component architecture — comprising the generator (G), the discriminator (D), and the auxiliary regressor (A) — was adopted because each component addresses a distinct requirement of the generation task. The generator receives both raw ECG waveforms and extracted numerical features as dual inputs, enabling it to leverage both the temporal morphology of the ECG signal and structured, clinically meaningful parameters (such as RR, PR, and QT intervals, wave-specific attributes, and heart rate variability measures) for waveform synthesis. The discriminator, formulated as a three-branch ID CNN that receives the patient's ECG data alongside the TDI waveform, was designed to optimize for patientspecific fidelity — ensuring that the generated waveform corresponds to the individual patient's electromechanical profile rather than producing a generic TDI morphology. The auxiliary regressor provides targeted feedback on the clinically most important waveform features — specifically the peak magnitudes of e', a', and s' velocities — thereby penalizing inaccuracies in the measurements that clinicians extract from TDI waveforms for diagnostic and prognostic decision-making.

[0066] In some examples, several alternative configurations were evaluated, including variations in which individual loss components were ablated (e.g., setting the auxiliary loss weight XA to zero), different optimizer selections for each model component, different augmentation strategies, different regularization techniques, and different network architectures. In particular, experiments using exclusively LI reconstruction loss without the adversarial loss component yielded outputs with diminished waveform features and, in many cases, unrealistic artifacts — most notably, the e' and a' diastolic waves exhibited severely underestimated amplitudes, and uncertainty in onset and timing caused the lowest point of the diastolic troughs to be stretched into horizontal plateaus (referred to herein as a "floor" artifact); in extreme cases, the e' and a' waves merged entirely, producing a floor artifact lasting over364937-9437-8896Docket No. (183161-37) 25% of the cardiac cycle. These failure modes demonstrated that the adversarial loss was useful to constrain the generator to produce single, realistic waveform outputs rather than blurred averages of all plausible waveforms. The specific hyperparameter values — ZR = 40, kG = 5, and kA = 2 for the loss function weights; Adam optimizer with learning rate IO-3for the generator, RMSprop with learning rate 3 x 10-4for the discriminator, and Adam with learning rate 5 x 10-3for the auxiliary model; 40 training epochs; and dropout rates of 50% for fully connected layers and 10% for convolutional layers — were selected through iterative trial-and-error evaluation, comparing validation reconstruction loss, correlations in key waveform intervals and peak magnitudes, loss curve behavior, and qualitative examination of individual generated waveform samples. The use of RMSprop with a comparatively lower learning rate for the discriminator, rather than Adam, was adopted to prevent the discriminator from overpowering the generator during early training, which would otherwise risk mode collapse. All training, hyperparameter tuning, and model development were restricted to data from a single site (US site #1) to ensure that performance on external validation datasets reflected true generalizability of the selected architecture and configuration.

[0067] At block 1204, method 1200 obtains and prepares a training dataset comprising paired ECG and echocardiography data. In some embodiments, the training dataset may comprise single or multi-lead ECG data over a period of time with corresponding echocardiography waveform data (e.g., TDI waveform data) acquired from the same patients, preferably contemporaneously or within a short temporal window (e.g., within the same clinical encounter or within a few hours of each other) so as to better ensure that the ECG and echocardiography data reflect substantially the same cardiac state. Preparation of the training dataset may include converting spectral TDI images from their original format (e.g., DICOM) into one-dimensional vectorized waveforms using an automated vectorization system guided by validated image landmarks; dividing each TDI recording into individual cardiac cycles and normalizing each374937-9437-8896Docket No. (183161-37) cycle to a standardized length; applying smoothing filters to reduce noise; dividing corresponding ECG signals into RR intervals and pairing each ECG cardiac cycle with a corresponding TDI cardiac cycle; and extracting numerical features from the ECG signals, such as basic intervals (RR, PR, QT), wave-specific attributes (axis, amplitude, duration), heart rate variability measures, morphological characteristics, dispersion measures, and corrected QT intervals across multiple leads.

[0068] In some embodiments, the extracted numerical features may optionally be normalized by subtracting the mean and scaling to unit variance as determined by samples in the training set.

[0069] Optionally, the training dataset may further include clinical labels associated with each patient, such as diagnoses of LV diastolic dysfunction, LV systolic dysfunction, ejection fraction values, or other clinical or echocardiographic classifications, for subsequent use in training downstream classifier or predictive models at block 1214. Or, in other embodiments, the training data sets may be labeled to include such diagnoses (whether by expert review, or via existing machine learning models for predicting such labels or classifications).

[0070] Optionally, at block 1206, an augmentation process may be applied to increase the number and variability of training samples. Such augmentation may include one or more of: selecting alternative ECG cardiac cycles from the same patient's recording for pairing with TDI cycles; phase-shifting aligned ECG-TDI pairs by a random percentage; setting random segments of the ECG signal to zero to simulate missing data; and removing random subsets of ECG leads by setting their values to zero. In some embodiments, the augmentation process may generate a plurality of additional samples (e.g., ten additional samples) for each original sample in the training set, with at least one augmentation technique applied to each additional sample. Augmentation of the training dataset in this manner can improve model robustness, including robustness to reduced-lead ECG inputs encountered at runtime.384937-9437-8896Docket No. (183161-37)

[0071] At block 1208, method 1200 may include training the GAN model using training dataset from block 1204. During training, the generator network may receive ECG waveform data and extracted ECG features as input and produce synthetic TDI waveforms. The discriminator network may receive both real TDI waveforms and synthetic TDI waveforms generated by the generator, along with the corresponding ECG data, and is trained to classify waveforms as real or synthetic, thereby providing adversarial feedback to the generator. The auxiliary regressor may be trained on both real and synthetic TDI waveforms to predict key waveform characteristics, such as the magnitudes of e', a', and s' velocities, and may optionally provide additional feedback to the generator by penalizing inaccuracies in these key features. The generator may be structured so as to minimize a total loss comprising one or more of: a reconstruction loss, an adversarial loss from the discriminator, and an auxiliary loss from the auxiliary regressor, wherein each may be weighted by respective hyperparameters. Training may proceed for a predetermined number of epochs, with the optimal model configuration selected based on validation metrics such as reconstruction loss, correlations in key waveform features, and / or qualitative examination of generated waveforms, hi embodiments where separate model instances are configured for different measurement locations (e.g., septal and lateral walls), each model instance is trained independently on the corresponding subset of the training data.

[0072] At block 1210, method 1200 validates the trained GAN model. Validation may be performed using held-out internal validation data (e.g., a subset of data from the same site used for training that was withheld from the training process) and / or external validation data such as from one or more independent clinical sites or patient examinations not used during training. Validation may include one or more of the validation approaches described below with respect to the inventors’ experiments and validation studies. If validation metrics do not meet predetermined acceptance criteria, method 1200 may return to block 1202 or block 1204 to394937-9437-8896Docket No. (183161-37) adjust model architecture, hyperparameters, training data, or augmentation strategies, and repeat training at block 1208.

[0073] At block 1212, method 1200 may optionally train or incorporate a foundation ECG model. In some embodiments, a foundation ECG model may be trained — or a pretrained foundation ECG model may be fine-tuned — using the same or overlapping ECG datasets used for GAN training at blocks 1204 and 1208, or using a larger corpus of ECG recordings. The foundation ECG model may be structured to receive raw multi-lead ECG signals as inputs and to generate high-dimensional embedding vectors representing latent features of the ECG signals. Unlike the clinically interpretable features extracted from synthetic TDI waveforms (e.g., e', s', a'), the embedding vectors generated by the foundation model comprise learned representations that may capture complex, non-linear patterns in the ECG signals not readily characterized by conventional clinical features. In some embodiments, the training datasets may be grouped by clinical designations (e.g., only ECG data from patients with a given cardiac condition; only ECG data from healthy patients with no cardiac condition, etc., so that the embeddings can be tailored). In some embodiments, the foundation ECG model may be pretrained in a self-supervised manner on a large corpus of ECG recordings (e.g., greater than one million recordings) and subsequently fine-tuned on the training dataset used for GAN development and / or condition-based subsets. The embedding vectors generated by the foundation ECG model may be used as additional input features for downstream classifier or predictive models trained at block 1214, either alone or in combination with features extracted from synthetic TDI waveforms generated by the trained GAN. Block 1212 is optional and may be omitted in embodiments where downstream classification is not used, or merely relies on other features such as features extracted from synthetic TDI waveforms and / or clinical data.

[0074] At block 1214, method 1200 may optionally train one or more classification or predictive models, using one or more of: TDI waveform data from the trained GAN; features404937-9437-8896Docket No. (183161-37) derived from outputs of the trained GAN model; features derived from actual TDI waveform data; and / or features determined by the foundation ECG model trained at block 1212. In some embodiments, the classification models receive as input one or more of: clinically interpretable features extracted from real TDI waveforms and / or synthetic TDI waveforms generated by the trained GAN (e.g., peak systolic velocity s', peak early diastolic velocity e', peak late diastolic velocity a', and interval-based timing features such as time to peak displacement, mechanical relaxation time, and the ratio of QT interval duration to time to peak displacement); clinical features associated with the patient (e.g., age, blood pressure, comorbidities), which may be obtained from the clinical labels optionally included in the training dataset at block 1204; conventional ECG interpretation data (e.g., automated ECG analysis program output); and / or high-dimensional embedding vectors generated by the foundation ECG model trained at block 1212. The classification models may be trained for specific diagnostic or prognostic tasks, such as predicting LV diastolic dysfunction, predicting LV systolic dysfunction (e.g., ejection fraction less than 50%), stratifying mortality risk, or predicting cardiovascular outcomes. In some embodiments, multiple candidate model types (e.g., naive Bayes, logistic regression, support vector machines, decision trees, random forests, k-nearest neighbors) may be trained and compared, with an optimal model selected based on performance metrics such as area under the receiver operating characteristic curve (AUC) and clinically acceptable sensitivity and specificity thresholds. Block 1214 is optional and may be omitted in embodiments where the desired output of method 1100 comprises the synthetic TDI waveform or extracted waveform parameters themselves, rather than a diagnostic or prognostic classification.

[0075] In some embodiments, at block 1214, one or more of the classification or predictive models may be trained to receive synthetic GLS as an input feature, either alone or in combination with other synthetic TDI-derived parameters (e.g., s', e', a', MAPSE), clinical features, and / or foundation ECG model embeddings. In such embodiments, the synthetic GLS414937-9437-8896Docket No. (183161-37) value is computed as described herein — by dividing the synthetic MAPSE by a predicted LV longitudinal length — and is provided as an additional input feature to the classification model. In some examples, the inclusion of synthetic GLS together with synthetic MAPSE and synthetic TDI velocities in a multivariate classification model improved discrimination for LV systolic dysfunction relative to models using synthetic TDI velocities alone, demonstrating the incremental diagnostic value of strain-based metrics derived from the synthetic TDI pipeline. In further embodiments, the machine learning model for predicting LV longitudinal length may itself be trained at block 1214 using the training dataset prepared at block 1204, wherein LV length measurements obtained from echocardiographic images in the training dataset serve as ground truth labels, and features extracted from the corresponding ECG signals and synthetic TDI waveforms serve as model inputs.

[0076] At block 1216, method 1200 packages and deploys the trained model(s) for use in runtime inference, such as in method 1100 of FIG. 13. Deployment may include packaging the trained GAN model (and optionally the foundation ECG model and / or downstream classifier models) into a software application or service configured for the target deployment environment — for example, a wearable device, a mobile application, a telemedicine platform, a cloud-based service accessible via communication network 130, or a clinical workstation. In some embodiments, the deployed model(s) may be configured to accept ECG input corresponding to the expected lead configuration of the target environment, including reduced-lead configurations where method 1000 of FIG. 12 is used to synthesize additional lead signals prior to model input. In further embodiments, deployment may include configuring the software application to present synthetic TDI waveforms, extracted parameters, and / or diagnostic or prognostic determinations to users via an output device, and to store or transmit results for longitudinal tracking, clinical decision support, or electronic medical record integration.424937-9437-8896Docket No. (183161-37)

[0077] Accordingly embodiments which adopt some or all of the aspects of such a method 1200 can achieve echocardiographic-equivalent data acquisition without: (1) the need for expensive, complex ultrasound equipment (instead, merely one or more electrodes for ECG-style cardiac electrical signal acquisition can be used); (2) the need for an ultrasound technician or other healthcare provider to perform the data acquisition in person; (3) inconsistency or manual error caused by inattentive or inexperienced operators (because, for example, once electrodes are affixed to a human’s skin / body, no further human skill is required); and (4) overuse of echocardiographic studies. Additionally, such methods provide improved ability for healthcare systems, radiology platforms, cardiology clinic medical record systems, and the like to generate assessments of biomechanical cardiac function with less data, and via longer monitoring periods. For example, a patient might wear electrodes affixed to their chest via adhesive patches for hours (or even days) at a time, while the constantly measure and monitor cardiac mechanical function and cardiac activity under real world scenarios of a variety of conditions.Example Hardware Integration of Hybrid Data Analysis System

[0078] FIG. 15 shows a block diagram illustrating a system 1300 for generating models, diagnostic information, and / or synthetic TDI waveforms and other echocardiogram data. A computing device 1302 can be mobile device, wearable device, cloud resource, medical monitoring equipment, tablet, workstation, or other computing device such as a single-board computer, a computing chip, or any suitable computing device that can receive sensor signals (e.g., single or multi-lead ECG signals) in real time or post-acquisition and generate clinically useful cardiac information for a given patient. Thus, the processes described below and in the present disclosure may be tied to training and running sensor data transformations and machine learning models for a specific “local” sensing device (e.g., ECG sensor). As further illustrated434937-9437-8896Docket No. (183161-37) in FIG. 15, the system 1300 includes a patient 1318 having one or more ECG electrodes or leads disposed on the patient's body surface for detecting cardiac electrical signals. The ECG electrodes or leads of the patient 1318 are communicatively coupled to an ECG machine 1320, which may comprise any suitable ECG acquisition device — including, without limitation, a clinical-grade multi-lead ECG system, a portable or handheld ECG monitor, a wearable ECG device, or a consumer-grade ECG-capable device — configured to detect, digitize, and / or transmit ECG signal data from the patient 1318. In some embodiments, the ECG machine 1320 may transmit acquired ECG signal data to the computing device 1302 via the communication network 1314 and / or via a direct wired or wireless connection through the input(s) 1312 of the computing device 1302, wherein the wireless connection may comprise any suitable short-range or long-range wireless communication protocol or combination thereof, including, without limitation, a Bluetooth connection (e.g., Bluetooth Classic or Bluetooth Low Energy (BLE)), a Wi-Fi connection (e.g., IEEE 802.1 la / b / g / n / ac / ax), a near-field communication (NFC) connection, a Zigbee connection, a Z-Wave connection, an ANT or ANT+ connection, a cellular connection (e.g., 3G, 4G LTE, 5G NR, or other cellular standard), an infrared (IR) connection, or any other suitable wireless communication link capable of transmitting ECG signal data from the ECG machine 1320 to the computing device 1302.

[0079] In the system 1300, a computing device 1302 can obtain or receive ECG signal information. The signal information can be sensor data received from an ECG detection system or directly from outputs of one or more ECG electrodes / leads. In some embodiments, computing device 1302 can also obtain or receive other (similar or complementary) sensor data and signals, such as motion data, photoplethysmography data, ballistocardiography data, force data, pressure data, temperature data, vibration or acoustic data, electrochemical sensor data, breath / respiratory sensor data, various electrical signal measurements, or the like. In other examples, the dataset can include one or more features or vectors extracted from the sensor444937-9437-8896Docket No. (183161-37) data or signals, or filtered or pre-processed data or signals. The computing device 1302 can receive the dataset, whether from a sensor or as stored in a database, via communication network 1314 and a communications system 1310 or an input 1312 of the computing device 1302.

[0080] In some embodiments, the processor 1304 can be any suitable hardware processor or combination of processors, such as a central processing unit (CPU), a graphics processing unit (GPU), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), a digital signal processor (DSP), a microcontroller (MCU), etc.

[0081] The memory 1306 can include any suitable storage device or devices that can be used to store suitable software, ML model(s), sensor data, and instructions that can be used, for example, by the processor 1304 to perform methods such as described herein. The memory 1306 can include a non-transitory computer-readable medium including any suitable volatile memory, non-volatile memory, storage, or any suitable combination thereof. For example, memory 1306 can include random access memory (RAM), read-only memory (ROM), electronically-erasable programmable read-only memory (EEPROM), one or more flash drives, removable media storage, one or more hard disks, one or more solid state drives, one or more optical drives, etc.

[0082] The computing device 1302 can further include a communications system 1310. The communications system 1310 can include any suitable hardware, firmware, and / or software for communicating information over a communication network 1314 and / or any other suitable communication networks. For example, the communications system 1310 can include one or more transceivers, one or more communication chips and / or chip sets, network cards, RF or cellular network adapters, etc. In a more particular example, the communications system 1310 can include hardware, firmware and / or software that can be used to establish a Wi-Fi connection, a Bluetooth connection, a cellular connection, an Ethernet connection, etc.454937-9437-8896Docket No. (183161-37)

[0083] The computing device 1302 can receive and / or transmit information (e.g., sensor data, training data, ground truth labels, software updates, ML models, classification algorithms, 3D modeling frameworks, additional patient data, a disease prediction indication, transformed data, output verification data, user preferences, a trained neural network, etc.) over a communication network 1314. In some examples, the communication network 1314 can be any suitable communication network or combination of communication networks. For example, the communication network 1314 can include a Wi-Fi network (which can include one or more wireless routers, one or more switches, etc.), a peer-to-peer network (e.g., a Bluetooth network), a cellular network (e.g., a 3G network, a 4G network, a 5G network, etc., complying with any suitable standard, such as CDMA, GSM, LTE, LTE Advanced, NR, etc.), a wired network, an EMR system, a DICOM system, encrypted wireless transmission, etc. In some embodiments, communication network 1314 can be a local area network, a wide area network, a public network (e.g., the Internet), a private or semi-private network (e.g., a corporate or university intranet), any other suitable type of network, or any suitable combination of networks. Communications links shown in FIG. 15 can each be any suitable communications link or combination of communications links, such as wired links, fiber optic links, Wi-Fi links, Bluetooth links, cellular links, etc.

[0084] In some examples, the computing device 1302 can further include one or more output devices 1308. The output device 1308 can include a data transmission module, such as to output a transformed dataset, 3D model, cardiac disease / condition inference, prediction or classification. In other examples, the output 1308 can include a display to present the foregoing to a patient, cardiologist, healthcare provider, surgical team, emergency team, nutritionist, therapist, or other user. In some embodiments, the output 1308 can include any suitable display devices, such as a computer monitor, a touchscreen, a television, an infotainment screen, etc. to display a report containing a classification or prediction, or a 3D visualization. In further464937-9437-8896Docket No. (183161-37) examples, the foregoing output types can be transmitted to another system or device over the communication network 1314. In further examples, the computing device 1302 can include an input 1312 using any suitable input devices (e.g., a keyboard, a mouse, a touchscreen, a microphone, etc.) and / or the one or more sensors that can produce the raw sensor data or the dataset 1316.Experimental Implementations and Validation Studies

[0085] The inventors conducted several studies and experiments validating the approaches described herein. These validation studies were conducted to demonstrate and quantify the advantages of the overall scope of embodiments contemplated. Although the studies were performed using certain implementations and configurations, it should be understood that the results of these studies establish the advantages of all contemplated embodiments. In other words, the particular examples used for purposes of the validation studies are not intended as limiting. One such experiment is described below, to serve as an illustrative example of the improvements achieved by embodiments of the present disclosure versus previous approaches.

[0086] Data acquisition and preprocessing

[0087] Across three patient sites, paired ECG and echocardiography images were obtained in the same tertiary care settings on the same day - within three hours of each other. Conventional 12-lead ECG was recorded and interpreted using the University of Glasgow Interpretive Analysis program (release 28.5, January 2014).

[0088] In some embodiments, comprehensive 2-dimensional (2D) Doppler echocardiography was performed using commercially available ultrasound equipment. Acquisitions were performed by licensed sonographers or doctors who were blinded to any other data of the present example. The echocardiography data was interpreted in a core lab for obtaining 474937-9437-8896Docket No. (183161-37) comprehensive 2D and Doppler measurements of LV chamber size and function, including TDI measurements of longitudinal mitral annular velocities from the septal and lateral comers during systole (s’), early (e’) and late (a’) diastole. LV diastolic dysfunction was defined per the 2016 ASE / EACVI guidelines with the following cutoff values suggesting abnormal diastolic function: 1) septal e' <7 cm / s and / or lateral e' <10 cm / s; 2) averaged E / e' >14; 3) tricuspid regurgitation (TR) velocity >2.8 m / s; and 4) left atrial volume index >34 ml / m2. Diastolic dysfunction was determined to be present when three or all of the above four parameters were abnormal. Left ventricle volumes were determined using the modified biplane Simpson's method to estimate EF. The left ventricular mass index was obtained as an indicator of LV hypertrophy (LVH) by echocardiography as a ratio of LV mass and body surface area. LVH was defined by LV mass index thresholds of 125 g / m2 for men and 110 g / m2 for women. A patient with LV systolic dysfunction (EF<50%) and LVH was also included to have diastolic dysfunction per current guideline recommendations, irrespective of the number of abnormal features.

[0089] ECG and tissue Doppler data preprocessing

[0090] A multi-step preprocessing pipeline was used to generate the final datasets used in training and evaluation. The 2D images of TDI waveforms required conversion into ID vectors to simplify, standardize, and eliminate redundant information. After conversion from the original DICOM format to a JPG image, an automated system guided by manually validated image landmarks was implemented for vectorization. The amplitudes of each timestep were determined by selecting the point that resulted in the greatest drop in pixel intensity when moving from the inner to outer envelopes. As the model is developed to generate the corresponding TDI from a single RR interval of an ECG, each TDI was split by cardiac cycle in this manner. This results in multiple samples being generated per patient. The length of each484937-9437-8896Docket No. (183161-37) cycle is then normalized to length 100, setting a standardized target size. A slight moving average smoothening is then applied to remove noise, with a filter size of 3 being used. Finally, the TDI cycles are rescaled by a constant. Corresponding 12-lead patient ECGs were divided into RR intervals, with unique cycles paired with each TDI cycle. This set of 12 waveforms was resized to length 512, with the raw ECG otherwise entirely preserved. In addition to their raw waveforms, an algorithm is implemented to extract 554 numerical features to include as input for the model. These include basic intervals (RR, PR, QT), wave-specific attributes (axis, amplitude, duration), and advanced vector measures. These features encompass detailed assessments of heart rate variability, morphological characteristics, dispersion measures, and corrected QT intervals across multiple leads. As a final step, the features are normalized by subtracting the mean and scaling to unit variance as determined by samples in the training set.

[0091] Finally, an augmentation step is applied to substantially increase the number of training samples. Ten additional samples are generated for each original sample in the training set. This is done in four ways: first, a new random cardiac cycle from the ECG is chosen, adding several new real cycles for training. Second, with a probability of 0.6, both the ECG and TDI have their phase shifted by a random percentage, remaining aligned. Third, with a probability of 0.5, a random segment of the ECG is set to zero, effectively imputing a time range. Fourth, with a probability of 0.5, 1-11 leads are removed, with every timestep set to zero. This process is configured to apply at least one augmentation step for each of the 10 augmented samples. Extracted ECG features are left unchanged in the augmentation process and are simply copied.

[0092] In this multi-step process, several patients and / or cycles were rejected. While typically occurring due to the simple issue of arbitrarily missing files, a minority is rejected due to the poor quality of either ECG or TDI. This prevents the samples from being processed owing to extreme noise or the inadequacy of TDI signals. For US site #1, this affected 23 samples in494937-9437-8896Docket No. (183161-37) lateral and 39 in septal, while there were only three samples for the Canada site for both lateral and septal.

[0093] GAN model development and training

[0094] A pair of conditional GANs inspired by the pix2pix architecture were trained to generate synthetic TDI waveforms from 12-lead ECGs, each tailored for lateral or septal views. The GANs consist of a generator (G) that produces TDI waveforms (TDIGAN) from ECG, a competing discriminator (D) that distinguishes real TDI waveforms (TDIreal) from synthetic, and an auxiliary regressor (A) that predicts key waveform characteristics (FIG. 2). The 12-lead ECGs are provided as both raw waveforms ECGwave and as features extracted in preprocessing, ECGfeats. In some embodiments, the ECG data is provided to a GAN or a trained generator network of the GAN.

[0095] The need for a learned loss function SGAN via D in addition to a simple LI reconstruction loss 3AA is primarily owed to variations in TDI waveforms for a single patient, including inter-beat differences and minor deviations in phase and period. With only SAA, G is optimized to generate a fuzzy average of all plausible TDI waveforms. Experiments using exclusively reconstruction loss yielded outputs with diminished features and, in many cases, unrealistic artifacts. This mainly affected the e’ and a’ waves, where amplitudes were often severely underestimated, and uncertainty in onset and timing resulted in the lowest point of the trough being stretched out into a horizontal “floor”. In extreme cases, the e’ and a’ waves could merge, resulting in a floor lasting over 25% of the cardiac cycle. Instead, employing an adversarial loss ^GAN optimized for a single, realistic output from the GAN that corresponds to the patient’s ECG. The relevant discriminator D was formulated as a three-branch ID Convolutional Neural Network (CNN) and was provided with the patient's ECGwave and ECGfeats to optimize for patient-specific features. D was trained with TDIreal and TDIGAN,504937-9437-8896Docket No. (183161-37)minimizing binary cross-entropy loss for classifying as real or synthetic. In practice, including a third loss SB via A, scaled by a small AA, slightly improved performance across several metrics by further penalizing inaccuracies in key points on the wave. A was a simple ID CNN regression model predicting the magnitudes of human-derived e', a', and waveform-derived s' from TDIreal. It was trained on both TDIGAN and TDIreal, minimizing mean squared error. Finally, G itself is a convolutional encoder-decoder network. Convolutional features extracted from ECGwave were concatenated with ECGfeats and transformed through fully connected layers before being upsampled with convolutional layers to generate TDIGAN. The final objective for the generator was to minimize the total loss. At the same time, the discriminator sought to maximize its component of the loss. This was formalized as:and being a vector of the target e’, a’, and s’ values.

[0096] For training, AR = 40, AG = 5, and AA = 2 were used. The optimizer configurations were tailored for different components of the model: the generator was optimized using Adaptive Moment Estimation (Adam) with a learning rate of 10-3, while the discriminator employed Root Mean Square Propagation (RMSprop) with a rate of 3 x 10-4, and the auxiliary model used Adam with a learning rate of 5 x 10-3. The models were trained for 40 epochs. All models used dropout for regularization at a rate of 50% for fully connected layers and 10% for514937-9437-8896Docket No. (183161-37) convolutional layers. During development, several choices of hyperparameters, optimizers, augmentation strategies, regularization techniques, and architectures were explored, including variations with different Xs set to 0. Ultimately, an optimal configuration was chosen via trial and error by comparing validation SR , measuring correlations in key intervals and a’, s’, e’ magnitudes, evaluating loss curves, and examining individual TDIGAN samples. All training, hyperparameter tuning, and other model development were restricted to data from US site #1.

[0097] Evaluation of synthetic waveforms

[0098] To evaluate the model's performance in predicting a complete waveform, the TDI GAN samples underwent a phase alignment step to match the corresponding TDI real samples. This alignment ensured that phase differences did not confound the performance evaluation. For patients with multiple available beats, alignment was performed on each beat, followed by averaging the aligned beats. Subsequently, regression and Bland-Altman analyses were conducted across all timesteps for all waveforms in the internal and external validation sets. These analyses assessed the correlation between synthetic and real TDI magnitudes throughout the cardiac cycle, with the degree of linear correlation serving as the primary metric. Finally, cosine similarity was computed for each pair of aligned synthetic and ground truth TDI waveforms to quantify their resemblance further.

[0099] Furthermore, the correlation between key time intervals in synthetic and real TDIs was examined, with a regression analysis performed for the following:

[0100] a) Time from the onset of the R wave on ECG to the peak of the ejection (s) wave on TDI524937-9437-8896Docket No. (183161-37)

[0101] b) The timing of mechanical contraction was developed from the onset of the R wave on ECG to peak displacement in the systole (developed as a time integral of the TDI velocity waveform).

[0102] c) Mechanical relaxation time was defined as the difference between the RR interval and the duration of mechanical contraction.

[0103] d) R-time to the onset of peak early diastolic velocity (e) wave on TDI.

[0104] e) The ratio between QT (reflecting electrical systole) and R to peak displacement (reflecting mechanical contraction).

[0105] With the TDIs normalized to one cardiac cycle, the known equivalent duration of the RR interval was used to determine, in milliseconds, the duration of each of the TDI-derived intervals.

[0106] Predicting left ventricular dysfimction

[0107] To demonstrate the diagnostic performance and incremental value of the synthetic TDI waveforms, a series of ML models, including logistic regression, support vector machines, decision trees, random forests, naive Bayes, and k-nearest neighbors, were developed for predicting LV diastolic and systolic dysfunction, and their performance was assessed and compared. An optimal model was then selected using the AUC and acceptable sensitivities. The ML models were developed using data from all patients within US site #1 (n=518, FIG.1), employing an 80-20 random split for training and internal validation, respectively. The models were then evaluated using data from all patients at Canada and US site #2 (n=585, FIG.1). An exception was made for models based on real TDI waveforms (and those compared against them), as these were only available for a subset of patients.534937-9437-8896Docket No. (183161-37)

[0108] First, a pair of naive Bayes models were developed as a baseline, the first for clinical features and the second for Glasgow ECG. Each had variants for LV diastolic and systolic dysfunction. Clinical features included age, systolic blood pressure, coronary artery disease, and hypertension. The Glasgow ECG interpretation was mapped to binary values as normal (0), borderline, or abnormal (1). These models were followed by the TDI-based models, taking advantage of simple amplitude features (septal wall e’, a’, s’, and lateral e’) as well as intervalbased features (time to peak displacement, time to peak velocity, QT interval duration, and the ratio of QT interval duration to time to peak displacement). To assess the incremental value of TDI features, a model using both Glasgow ECG and clinical features was compared with one utilizing Glasgow ECG, clinical, and TDI-derived features. Across all models, missing features were handled by imputing the mean of the feature across all patients in the training set.

[0109] Along with evaluating the predictive performance and incremental value of synthetic waveforms versus Glasgow ECG and / or clinical features, a comparative analysis was performed to assess the performance difference, if any, between synthetic and real TDI waveforms. Due to the absence or unreadability of the real TDIs, the patient set used for training and evaluation differed from that of the GAN-based naive Bayes. To address this discrepancy, a secondary GAN-based model was developed solely for comparison, restricted to the same patient set as real for both training and evaluation.

[0110] To generate various threshold-dependent metrics (such as sensitivity and specificity), thresholds were automatically chosen via the internal validation set using Youden’s index to optimize performance for each model. This was done for all feature sets, except for those that exclusively used Glasgow ECG - as a binary feature, there are only two possible prediction probabilities. They were thus directly mapped to their respective classes.[OHl] External ECG dataset544937-9437-8896Docket No. (183161-37)

[0112] The present example utilized a publicly available annotated dataset of 12-lead ECGs. This included individual patient files containing raw 12-lead ECG tracings, along with an 'exam.csv' file that provided corresponding clinical data and follow-up survival information. The raw ECG tracings were processed to isolate individual cardiac cycles, which were then resized to a uniform length of 512 timesteps. For each patient, four cardiac cycles were randomly selected and provided, in some embodiments, to a GAN or a trained generator network of a GAN to generate four TDI cycles. These generated cycles were averaged to produce a single representative result. This procedure was carried out in parallel for multiple patients, facilitating analysis across the entire dataset.

[0113] Alterative Practical Application: Development of Atrial fibrillation-Tuned Model

[0114] To further demonstrate the ability of the model to learn associations unique to atrial fibrillation, additional augmentation was performed to the original training dataset, where, in half of all samples, the a’ deflection was ablated by replacement with the baseline velocity. Correspondingly in the associated ECG cycles, the interval from the end of the T wave to the end of the P-wave was replaced with those from randomly chosen patients with atrial fibrillation in an external arrhythmia-focused ECG dataset.

[0115] Study Participants and Clinical Characteristics

[0116] In one example, the present disclosure presents a post hoc analysis of a multicenter study that analyzed paired ECG and echocardiographic data from 1,103 subjects prospectively enrolled across three North American sites for developing machine learning models to predict diastolic dysfunction from surface ECG (FIG. 1). The detailed inclusion and exclusion criteria are listed as supplementary information. The investigation included analysis in two steps (FIG.1):554937-9437-8896Docket No. (183161-37)

[0117] a) The development and validation of synthetic TDI waveforms: The synthetic TDI model was developed from 9,144 lateral and 8,722 septal TDI-ECG pairs. For this, DICOM copies were made of pulsed Doppler TDI waveform images from a subset of patients (n=463) from US site #1. Pulsed spectral TDI waveforms were converted to single-line waveforms using a computerized algorithm. During this step, inadequate quality or noisy TDI data were excluded (n=26 for lateral, n=42 for septal). Since each patient’s TDI recordings were performed over multiple heartbeats, derivation of 1,014 lateral and 962 septal distinct TDI waveforms over a cardiac cycle. During prospective ECG recordings, each patient’s 12-lead ECG was recorded across several (>10) cardiac cycles. In the training set, each TDI cycle was crossmatched with multiple ECG cycles recorded for each patient and applied augmentations for further diversity. This resulted in an expanded final dataset of 9,144 lateral and 8,722 septal TDI-ECG pairs, of which 8,943 lateral and 8,536 septal TDI-ECG cycle pairs were used for training, and 201 lateral and 186 septal TDI-ECG pairs were used for internal validation. Following a similar approach, external validation was performed on the Canada site, using 816 lateral and 869 septal ECG-TDI beats (311 patients).

[0118] b) Validation of synthetic TDI waveforms using human-based measurements: While the GAN model training involved the conversion of raw spectral pulsed wave TDI recording into a single line TDI with automated measurements, the present disclosure also compares the synthetic waveform measurements with clinical measurements recorded directly from the pulsed wave TDI. Human expert echocardiography measurements were performed across training (518 participants at US site #1) and two external validation sites (585 patients from US site #2 and Canadian site), and the associations of the real and synthetic TDI measurements with clinical variables were compared. Clinical models were developed using synthetic TDI waveform measurements to predict diastolic and systolic dysfunction.564937-9437-8896Docket No. (183161-37)

[0119] Development and Validation of GAN-generated tissue Dowler -waveforms

[0120] The conditional GAN model combines a generator and discriminator assembly (FIG.2) with an auxiliary regressor providing additional feedback. Initially, the 12-lead ECG signal and its features were passed to the generator stage, where TDI waveforms were produced from the ECG signals. These generated waveforms were then distinguished as real or synthetic using the discriminator stage, providing feedback to the generator. Further refinement was achieved via feedback from the auxiliary.

[0121] Like the real TDI waveform, the synthetic TDI signal showed three peaks during a cardiac cycle: a positive systolic peak and two negative diastolic peaks (FIG. 3A). The positive systolic wave (s’ velocity) represents myocardial contraction. The negative waves represent the early diastolic myocardial relaxation (e’ velocity) and active atrial contraction in late diastole (a’ velocity). The average magnitude profiles of the real and synthetic TDI waveforms across all patients at septal and lateral walls also show the three waveforms (FIG. 3B).

[0122] Quantitative evaluations assessed the similarity between synthetic and real TDI waveforms in the external validation set (FIGS. 3C-3D. Each pair of line waveforms (real and synthetic) was compared at every 1% timestep of the cardiac cycle. Linear regression (R2) captured the amplitude correlation at each timestep, and cosine similarity (ranging from -1 to 1) quantified overall shape similarity. Bland- Altman analysis highlighted the absolute differences between the synthetic and real waveforms. Since guidelines recommend averaging TDI velocities from septal and lateral walls, measurements were analyzed individually for septal and lateral walls and averaged. On comparing the entire waveform, the synthetic waveform measurements averaged from septal and lateral walls exhibited a high cosine similarity (0.89, p<0.0001) and R2 (0.79, p<0.0001), demonstrating consistent similarity with real waveforms across all sets, maintaining robustness across an unseen site (FIG. 3C). The574937-9437-8896Docket No. (183161-37) Bland-Altman analysis of the averaged waveforms (FIG. 3D) revealed a minor mean bias of - 0.23 cm / s with the limits of agreement, ranging from -3.94 cm / s to 3.48 cm / s.

[0123] Randomized and controlled visual Turing Test of GAN-based Synthetic TDI -waveforms

[0124] To determine the realism of GAN-generated TDI waveforms as perceived by experts, a controlled, randomized visual Turing test was conducted on the external validation set with 80 TDI waveform cycles — 40 predicted by a GAN-based model and 40 real — assessed by four board-certified echocardiographers. All 80 TDI waveforms were shuffled and then presented to the echocardiographers. They were tasked with evaluating each waveform as either "real" or "synthetic" and labeling each as "normal" or "abnormal". The four echocardiographers achieved an average correct answer rate of 40% for distinguishing "real" or "synthetic" waveforms, lower than the expected 50% for random guesses (p = 0.20 by z-test for two proportions). Furthermore, averaged reader accuracies, sensitivities, and specificities were below 50%. This consistently demonstrates that GAN-generated and real TDI waveforms are indistinguishable. Furthermore, when tasked with identifying each waveform as normal or abnormal, the echocardiographers showed strong agreement in their classifications (averaged normal = 32 and abnormal = 48), treating synthetic and real TDI waveforms similarly (K = 0.80).

[0125] Comparison of GAN-generated TDI -waveforms with clinical and echocardiography features

[0126] Amongst the different TDI waveforms, early diastolic velocity measurements (e’) have been specifically recommended in guidelines for routine assessment of LV diastolic function. Human expert-based measurements were performed from lateral and septal walls for the spectral pulsed Doppler TDI waves, blinded to the synthetic measurements. Average e' values from GAN-generated and real TDI waveforms were compared against clinical parameters such 584937-9437-8896Docket No. (183161-37) as age, systolic blood pressure, and diastolic blood pressure. The average e’ values measured from both GAN-generated and real TDI waveforms exhibited similar associations, as evidenced by significant R-values (p<0.0001) when compared against age, systolic blood pressure, and diastolic blood pressure. Furthermore, univariate logistic regression analyses were performed between clinical and echocardiography features and risk factors (Fig. 4). A 1 cm / s increment in average e’ measured from synthetic TDI waveform reduced the odds of association with clinical features such as age [>65 years; Odds ratio (OR)=0.74, 95% CI: 0.68-0.80], diabetes mellitus (OR=0.86, 95% CI: 0.79-0.94), hypertension (OR=0.77, 95% CI: 0.71-0.83), dyslipidemia (OR=0.87, 95% CI: 0.81-0.94) and with echocardiographic features such as pulmonary hypertension (OR=0.82, 95% CI: 0.75-0.89), reduced EF (OR=0.81, 95% CI: 0.71-0.93), LV hypertrophy (OR=0.72, 95% CI: 0.64-0.80), LV diastolic dysfunction (OR=0.78, 95% CI: 0.72-0.84), or valvular heart disease (OR=0.84, 95% CI: 0.77-0.93). The odds ratio for real and synthetic e’ (FIGS. 4A and 4B respectively) showed striking similarity, providing additional validation that the model, beyond learning to generate the shape, also varied in an individual patient, depending on their unique risk profile.

[0127] Diagnostic value of GAN-based tissue Doppler waveforms

[0128] Since TDI detects early signs of LV dysfunction, one validation study relating to an embodiment of the present disclosure assessed whether simple magnitude and timing measurements of the synthetic TDI waveforms could also be used as a screening tool to optimize the referral of patients for echocardiographic LV function assessment. This would allow the ECG model to be converted to a mechanical signal from which simple features are extracted for clinical decision-making. For screening and referral, the present disclosure assessed the incremental value of the synthetic TDI measurements over clinical data and conventional ECG assessment. This validation study used the automated ECG analysis594937-9437-8896Docket No. (183161-37) program from the University of Glasgow (release 28.5, issued in January 2014), a well-established computer software for automated ECG measurements and abnormality detection. This program has been extensively evaluated, is used worldwide, and is well standardized to detect an abnormal echocardiogram based on all the standard amplitude, duration, axis measurements, rhythm analysis, and diagnostic interpretations.

[0129] To keep the screening prediction models interpretable and straightforward, TDI magnitude and durations were combined into several simple tabular machine learning models. The naive Bayes model, based on the Bayes theorem, with the additional ‘naive’ assumption that features are conditionally independent, stood out as the best approach. In the external validation set, 157 (26.8%) patients had echocardiographic diagnoses of LV diastolic dysfunction using criteria stipulated in the 2016 ASE / EACVI recommendations. The Naive Bayes model developed using features from the synthetic TDI showed an AUC of 0.77 (95% confidence interval, CI: 0.73-0.81) in comparison with a similar model developed using clinical and Glasgow ECG analysis, which showed AUCs of 0.72 (95% CI: 0.67-0.76) and 0.67 (95% CI: 0.63-0.71), respectively on external validation. Furthermore, when combined with the ECG and clinical features, the synthetic TDI demonstrated an AUC of 0.80 (95% CI: 0.75-0.84), showing significant incremental value (FIG. 5A, p<0.001) over a model using the ECG and clinical features (AUC of 0.74, 95% CI: 0.70-0.79). Additionally, on substituting the synthetic TDI features with those from a real TDI in the Canadian cohort, a similar model performance was observed (AUC of 0.82, 95% CI: 0.77-0.86 vs. 0.81, 95% CI: 0.76-0.86, p=0.82, respectively).

[0130] Similarly, the inventors’ validation study also examined the ability to identify patients with systolic dysfunction (EF<50%). 49 (8.4%) patients had EF<50% in the external validation set. The naive Bayes model developed using the synthetic TDI was superior with an AUC of604937-9437-8896Docket No. (183161-37) 0.79 (95% CI: 0.72-0.85) compared to a similar model developed using clinical and Glasgow ECG analysis, which showed AUCs of 0.66 (95% CI: 0.59-0.72) and 0.72 (95% CI: 0.67-0.75) respectively on external validation (p<0.001 and p<0.05, respectively). The model combining ECG, clinical features, and synthetic TDI demonstrated an AUC of 0.81 (95% CI: 0.75-0.86). The inclusion of clinical and ECG data to TDI features showed a significant improvement (p<0.001) compared to a model using synthetic TDI features alone and showed better performance (FIG. 5B) than the model using the ECG and clinical features (AUC of 0.73, 95% CI: 0.66-0.79, p<0.001). Additionally, the synthetic TDI features were substituted with those from a real TDI in the Canadian Cohort, where both synthetic and real single-line waveforms showed similar model performance (AUC of 0.78, 95% CI: 0.74-0.82 vs. 0.80, 95% CI: 0.75-0.84, p=0.60, respectively).

[0131] The impact of the synthetic TDI model was assessed over the clinical features and Glasgow ECG analysis to improve the selection of patient referrals for echocardiograms. Using a synthetic TDI-based model to select people for echocardiography could reduce the number of echocardiograms by 64.3% and 69.9% [(true negatives + false negatives) / total screened x 100] for detecting LV systolic and diastolic dysfunction. However, 2.1% and 10.6% of cases with systolic or diastolic dysfunction would be missed. Alternatively, the Al model incorporating clinical features, abnormal Glasgow ECG, and synthetic TDI would result in a 60% and 58.3% reduction in echocardiograms but would miss 1.4% and 6.5% of LV systolic and diastolic dysfunction cases.

[0132] Prognostic value of GAN-based tissue Doppler waveforms

[0133] The validation study also assessed the prognostic value of synthetic TDI waveforms in a dataset of 12-lead ECGs including 233, 647 patients (51-20 years, 41% male) collected by the Telehealth Network of Minas Gerais, Brazil in the period between 2010 and 2016. The614937-9437-8896Docket No. (183161-37)baseline characteristic of the cohort has been previously published. During a median followup period of 3.5 (interquartile range, 2.1-5.2 years), 8341 (3.6%) died. The ECG-derived synthetic TDI were successfully batch predicted from ECG recordings in 228, 226 (98 % cases). The synthetic average e’ and s’ were associated with overall mortality (Table 1), even after multivariable adjustment for age, sex, the presence of atrial fibrillation, and the underlying pattern of conduction during sinus rhythm (adjusted hazard ratios, 0.97 (95% CI, 0.96-0.98), p<0.0001 and 0.91 (95% CI, 0.90-0.93), p<0.0001, for e’ and s’ respectively).

[0134] Table 1 Univariate and multivariate analysis for predicting mortality.Feature Hazard Ratio (95% CI) p-value Hazard Ratio (95% CI) p-value Age (yrs) 1.07 (1.06-1.07) <0.0001 1.06 (1.06-1.06) <0.0001 Male sex 1.75 (1.67-1.83) <0.0001 1.65 (1.58-1.73) <0.0001 AF 5.66 (5.21-6.15) <0.0001 2.08 (1.91-2.27) <0.0001 Sinus - 1°AVB 2.61 (2.32-2.93) <0.0001 1.25 (1.11-1.41) <0.0001 Sinus - RBBB 2.35 (2.14-2.58) <0.0001 1.15 (1.05-1.27) 0.004 Sinus - LBBB 3.47 (3.13-3.85) <0.0001 1.41 (1.27-1.57) <0.0001 Sinus - Bradycardia 0.59 (0.46-0.75) <0.0001 0.54 (0.42-0.69) <0.0001 Sinus - Tachycardia 2.28 (2.06-2.53) <0.0001 2.30 (2.07-2.56) <0.0001 Average e' (cm / sec) 0.85 (0.84-0.86) <0.0001 0.98 (0.96-0.99) <0.0001 Average s' (cm / sec) 0.77 (0.75-0.78) <0.0001 0.92 (0.90-0.94) <0.0001 AF = atrial fibrillation, 1°AVB = Ist-degree atrioventricular block, RBBB = right bundle branch block, LBBB = left bundle branch block

[0135] FIG. 6 shows the Kaplan-Meier curves depicting the cumulative probability of survival for persons stratified into synthetic TDI-derived average e’ tertiles, depicted separately for patients in sinus rhythm and atrial fibrillation. The relative mortality risk increased with increasing e’ tertiles (P<0.01); thus, persons with an e’ index in the third tertile had 12% higher risk of death compared to those with an e’ index in the first tertile (hazard ratio 1.12, 95% CI,624937-9437-8896Docket No. (183161-37)1.01 to 1.23; P=0.02). Similarly, for patients with atrial fibrillation, persons with an e’ index in the third tertile had 68% higher risk of death compared to those with an e’ index in the first tertile (hazard ratio 1.68; 95% CI, 1.63 to 1.74; PO.OOOl).

[0136] Synthetic average s’ and e’ predicted death even in patients with normal ECG (interpreted by the University of Glasgow ECG analysis program and expert review as previously described) (hazard ratio, 0.85; 95% CI, 0.82 to 0.88; PO.OOOl and 0.89; 95% CI, 0.87 to 0.91; PO.OOOl for s’ and e’, respectively). Moreover, for the entire cohort with sinus rhythm and atrial fibrillation, a model containing ECG, average e’ and average s’ showed incremental value over ECG alone for predicting death (c-statistic, 0.62, 95% CI 0.62-0.63 versus 0.69, 95% CI 0.68-0.69, pO.0001).Alternate Practical Application - Embedding Synthetic Echocardiographic Motion into ECG Models for Early and Scalable Cardiac Disease Detection

[0137] This example presents modeling framework that embeds synthetic echocardiographic motion — derived from 12-lead ECG — with echo-derived DD risk phenotypes to enable scalable detection of cardiac structural and functional abnormalities.

[0138] In one example, echocardiography derived nonlinear, geometry-informed DD risk phenotypes were used to train an ensemble ECG model combining: (1) features from synthetic cardiac motion generated from 12-lead ECGs via generative adversarial network, and (2) embeddings from a foundation ECG model pretrained on >10 million recordings. Development used a multicenter ECG-echo cohort (n=l,012), with validation in an independent test cohort (n=956), EchoNext (n=100,000; Columbia), and CODE-15% (n=233,770; Brazil with longitudinal outcomes) (FIGS. 7A and 7B).

[0139] The ensemble model accurately classified derived DD risk phenotypes (FIG. 7A, AUC= 0.86 [95% CI: 0.83-0.89] development; 0.85 [95% CI: 0.83-0.88] external test) with634937-9437-8896Docket No. (183161-37) incremental value seen over ECG foundation models (Net Reclassification Improvement=0.54, p<0.05). High-risk ECG phenotypes predicted structural remodeling (increased LV mass index, left atrial volumes, E / e'; reduced e') and associated strongly with clinical risk factors (age, hypertension, chronic kidney disease; all p<0.0001). In EchoNext, the model identified structural heart diseases with high discrimination (AUC 0.74-0.83 for aortic stenosis, valvular disease, ventricular dysfunction). In CODE- 15%, it predicted heart failure-related death (FIG.7B, Gray's test p=0.0002), with high-risk patients exhibiting three-fold higher 4-year mortality (8.5% vs 3%).

[0140] The present example illustrates that using echo-derived risk states and synthetic motion in ECG models may improve early detection of cardiac structural and functional abnormalities and optimize echocardiography use in high-risk, low-resource settings.

[0141] In another example, data from a primary multicenter cohort and two external validation datasets were used. The primary cohort consisted of 1,968 patients recruited from four North American sites. The primary cohort was divided into training (n=l,012) and external validation (n=956) sets. Two publicly available datasets were utilized for additional validation: EchoNext (100,000 ECGs from Columbia University with paired echocardiographic labels) and CODE-15% (345,779 ECG examinations from 233,770 patients from Brazil with mortality data).

[0142] Data Acquisition and Preprocessing

[0143] For the primary cohort, standard 12-lead ECGs (500 Hz) were paired with comprehensive 2DDoppler echocardiography, with both modalities obtained on the same day, preferably within three hours. ECGs were interpreted using the University of Glasgow Interpretive Analysis program (release 28.5, January 2014). Tissue Doppler imaging captured longitudinal myocardial velocities at the lateral and septal mitral annulus, with all echocardiographic parameters measured according to American Society of Echocardiography 644937-9437-8896Docket No. (183161-37) guidelines. All ECG recordings underwent band-pass filtering (0.5-40 Hz) and quality control. ECG waveforms were segmented from R-peak to R-peak, with each cardiac cycle normalized to 512 samples through linear interpolation and smoothed using a moving average filter (window size = 3). Cardiac velocity waveforms were similarly segmented and normalized to 100 time points per cycle. CODE-15% ECGs (400 Hz) and EchoNext ECGs (500 Hz) underwent identical preprocessing.

[0144] Generating Synthetic Tissue Motion

[0145] Synthetic cardiac motion waveforms were generated from 12-lead ECGs using a generative adversarial network (GAN) trained on 17,866 Echo-ECG pairs. This model generates synthetic cardiac velocity and displacement waveforms at 1% intervals across the cardiac cycle. External validation demonstrated strong correlation (r = 0.90, P < 0.0001) and cosine similarity (0.89, P <0.0001) between synthetic and real waveforms, with board-certified echocardiographers unable to distinguish between them in visual Turing testing.

[0146] Feature Engineering and Model Development

[0147] Comprehensive features were extracted from synthetic velocity and displacement waveforms across four domains: time-domain (21 features per waveform), wavelet-based from 3-levelDaubechies 4 decomposition (49 per waveform), peak morphology (12 per velocity waveform for S', e', a' peaks; 4 per displacement waveform), and frequency-domain from Fast Fourier Transform (11 per waveform). Eight direct peak amplitude measurements were included. From364 total features, the Boruta algorithm selected 34 features including age and sex for final model development. Following standardization, machine learning models were developed (FIG. 8) using H2O.aiAutoML with 10-fold cross-validation. Three models were developed: Model 1 utilized the 34cardiac velocity-based features; Model 2 employed recursive feature elimination to select the top 100 features from 1024-dimensional embeddings 654937-9437-8896Docket No. (183161-37) generated by a pre-trained ECG foundation model trained on over 10 million recordings; and an Ensemble Model combined predictions from Models 1 and 2 using stacking to generate a combined risk score reflecting the patient's overall cardiac function assessment.

[0148] In some examples, of 1,012 patients in the development cohort, 208 (20.5%) were high-risk and 804 (79.5%) low-risk. High-risk patients were older (median 68 vs 55 years, p<0.0001), predominantly male(62.5% vs 46.0%, pO.OOOl), with higher systolic blood pressure (133 vs 126 mmHg, p<0.0001).Traditional cardiovascular risk factors were more prevalent, including diabetes mellitus (34.3% vsl7.2%, p<0.0001), hypertension (85.8% vs 53.2%, p<0.0001), hyperlipidemia (78.4% vs 54.2%, pO.OOOl), and chronic kidney disease (6.7% vs 2.5%, p=0.01).High-risk patients demonstrated impaired systolic function (ejection fraction 54% vs 65%,p<0.0001) and characteristic diastolic dysfunction features: lower E / A ratio (0.83 vs 1.00, pO.OOOl), reduced e' velocity (6.64 vs 9.44 cm / s, pO.OOOl), and elevated E / e' ratio (11.23 vs7.51, pO.OOOl). Structural remodeling was evident with increased left ventricular mass index(99.00 vs 72.82 g / m2, pO.OOOl), left atrial volume index (30.70 vs 21.95 mL / m2, pO.OOOl), and tricuspid regurgitation velocity (2.60 vs 2.24 m / s, pO.OOOl).

[0149] Model performance

[0150] The ensemble model demonstrated robust discriminative performance for DD (FIG.7A). In the training cohort, AUC was 0.86 (95% CI: 0.83-0.89) with 77.4% sensitivity, 79.9% specificity, and 79.4% accuracy. Performance remained robust in external testing: AUC = 0.85 (95% CI: 0.83-0.88), 80.9% sensitivity, 74.5% specificity, and 76.9% accuracy. The ensemble model outperformed the ECG foundation model alone (FIG. 9A): AUC = 0.85versus 0.835 (p<0.05). Net reclassification analysis showed total NRI of 0.54 (95% CI 0.41-0.65, pO.OOOl), with NRI for non-events of 0.69. The ensemble correctly maintained 590 patients at low / intermediate risk, appropriately reclassified 94 patients (44 upgraded, 50 downgraded),664937-9437-8896Docket No. (183161-37)and identified 272 patients as high risk, showing improved stratification with minimal misclassification. The model's high-risk predictions demonstrated significant associations with established clinical parameters in both training and test sets (FIG. 9B). Age >65 years showed the strongest association in both the training set (OR: 1.05, 95% CI 1.04-1.05) and test set (OR: 1.06, 95% CI1.05-1.06). Other significant associations included elevated E / e', increased LVMi and LAVi, reduced e', hypertension, diabetes mellitus, and chronic kidney disease, with consistent effect sizes across both datasets.

[0151] Diagnostic Performance on EchoNext Dataset

[0152] External validation (n=35,718) demonstrated identification of diverse cardiovascular conditions (FIG. 9C). Highest discrimination was achieved for aortic stenosis detection (AUC=0.83, 95%CI: 0.81-0.84), reduced LVEF <50% (AUC=0.77, 95% CI: 0.77-0.78), RV systolic dysfunction(AUC=0.74), LV hypertrophy (AUC=0.73), mitral regurgitation (AUC=0.72) and aortic regurgitation (AUC=0.72), including TR velocity >3.2 m / s (AUC=0.72), PASP >45 mmHg(AUC=0.69) and the composite of all structural heart disease (AUC=0.74, 95% CI: 0.74-0.74).Prognostic Performance on CODE-15% Dataset External validation on CODE-15% demonstrated prognostic utility for HF-related mortality(Gray's test P=0.0002, SHR=4.27, 95% CI: 1.91-9.56, P=0.0004, FIG. 7B). By 4 years, high-risk patients reached 8.5% cumulative incidence versus 3% in low-risk patients, with early divergence within 2 years suggesting actionable prognostic information.Alternative Practical Application: Surface Electrocardiogram-Based Digital Twin of Left Ventricular Mechanics

[0153] This example presents using an electrocardiogram (ECG)-based cardiac digital twin that continuously maps surface electrical activity to LV mechanics, generating interpretable674937-9437-8896Docket No. (183161-37)markers — including tissue Doppler imaging (TDI) velocities, mitral annular plane systolic excursion (MAPSE), and global longitudinal strain (GLS).

[0154] In one example, data were used from 1,410 patients: 729 hospitalized with acute myocardial infarction and 681 age-matched outpatients with or without chronic LV dysfunction. An artificial intelligence pipeline synthesized TDI digital twin waveforms from 12-lead ECGs and estimated MAPSE and GLS.

[0155] Across both cohorts, synthetic LV metrics showed directionally consistent associations with clinical characteristics and mirrored trends observed in real TDI measurements (FIG. 10). LV systolic dysfunction was present in 372 patients (26%), including 103 outpatients with chronic dysfunction not attributed to acute myocardial infarction. In the full cohort, a multivariate model using synthetic LV mechanics achieved an area under receiver operator characteristic curve of 0.77 [95% CI 0.74-0.79, sensitivity 64%, specificity 77%] and 0.78 [95% CI 0.75-0.80, sensitivity 76%, specificity 68%] for predicting ejection fraction <50% and <30%, respectively.

[0156] This example presents that synthetic LV biomechanics present an interpretable approach for predicting LV systolic dysfunction.Alternative Practical Application: Synthetic Left Ventricular Mechanics Derived from Surface Electrocardiogram Predict Cardiac Mortality

[0157] The example of the present disclosure presents using synthetic tissue Doppler imaging (TDI)-derived left ventricular mechanics from surface ECG to evaluate scalable mortality risk stratification in acute myocardial infarction (AMI).684937-9437-8896Docket No. (183161-37)

[0158] In one example, 219,564 ECGs from a Brazilian population cohort and 729 AMI patients from a U.S. clinical cohort were used as data. Using a generative Al framework, tissue Doppler imaging (TDI) waveforms were synthesized from surface ECGs to extract systolic (s'), early diastolic (e'), late diastolic (a') velocities, and mitral annular plane systolic excursion (MAPSE). Cox models were used to assess associations with mortality in both cohorts.

[0159] In the population cohort, 817 AMI deaths (0.37%) occurred over a median 3.5-year follow-up. While all synthetic TDI metrics predicted AMI death (p < 0.001; FIG. 11), MAPSE was the strongest multivariate predictor, with each 1-mm increase associated with markedly lower mortality (HR: 0.19; 95% CI: 0.13-0.26; p < 0.001). Adding synthetic TDI metrics to ECG improved discrimination for AMI mortality (C-index: 0.70 vs. 0.61; p < 0.001). Similar associations were seen in the US clinical cohort (FIG. 11), where the model retained prognostic value for all-cause mortality (C-index: 0.64; 95% CI: 0.57-0.70).694937-9437-8896

Claims

1. Docket No. (183161-37)CLAIMSWHAT IS CLAIMED IS:

1. A system for generating synthetic echocardiographic waveform data from electrocardiogram (ECG) signals, the system comprising:at least one ECG electrode configured to be affixed to a surface of a patient and to detect electrical cardiac signals therefrom;a processor;an electronic data connection between the at least one ECG electrode and the processor; an output connection coupled to the processor and configured to convey output data to an output device; anda computer-readable memory coupled to the processor and storing instructions that, when executed by the processor, cause the processor to:receive, via the electronic data connection, the electrical cardiac signals detected by the at least one ECG electrode;preprocess the electrical cardiac signals to generate ECG data comprising at least one segmented cardiac cycle;provide the ECG data as input to a trained generator network, wherein the trained generator network was trained on a training dataset comprising paired ECG data and real tissue Doppler imaging (TDI) waveforms acquired from echocardiographic imaging of a plurality of patients;obtain, from the trained generator network, synthetic echocardiographic waveform data temporally-correlated with actual cardiac tissue movement of the patient over the at least one segmented cardiac cycle; andprovide the temporally-correlated synthetic echocardiographic waveform data in human-interpretable format to the output device via the output connection.

2. The system of claim 1, wherein the temporally-correlated synthetic echocardiographic waveform data comprises a synthetic tissue Doppler imaging (TDI) waveform temporally-correlated with myocardial tissue velocity of the patient during the at least one segmented cardiac cycle.

3. The system of claim 2, wherein the trained generator network was trained as a component of a generative adversarial network (GAN) further comprising: (i) a discriminator704937-9437-8896Docket No. (183161-37)network trained to distinguish real TDI waveforms from synthetic TDI waveforms generated by the trained generator network; and (ii) an auxiliary regressor trained to predict peak velocity characteristics of TDI waveforms, wherein the discriminator network and the auxiliary regressor each provided training feedback to the trained generator network during training of the GAN.

4. The system of claim 2, wherein the synthetic TDI waveform characterizes a three-peak morphology of the cardiac cycle comprising a systolic velocity peak (s'), an early diastolic velocity peak (e'), and a late diastolic velocity peak (a').

5. The system of claim 4, wherein the instructions further cause the processor to: derive synthetic average e' and s' values from the synthetic TDI waveform; and generate a mortality risk prediction for the patient based at least in part on the synthetic average e' and s' values, adjusted for one or more of age, sex, presence of atrial fibrillation, and cardiac conduction pattern.

6. The system of claim 1, wherein the human-interpretable format comprises: a visual depiction of the synthetic echocardiographic waveform data; and a graphical indication of a basis for a cardiac state determination corresponding to a clinical diagnostic assessment.

7. The system of claim 1, wherein the instructions further cause the processor to: extract magnitude and timing measurements from the synthetic echocardiographic waveform data; and provide the extracted magnitude and timing measurements as input to a trained machine learning classification model configured to generate a cardiac screening determination based thereon.

8. The system of claim 1, wherein the synthetic echocardiographic waveform data exhibits a cosine similarity of at least 0.75 with corresponding real echocardiographic waveform data acquired from echocardiographic imaging of the patient.

9. The system of claim 1, wherein the synthetic echocardiographic waveform data exhibits a linear correlation coefficient (R2) of at least 0.70 with corresponding real echocardiographic waveform data acquired from echocardiographic imaging of the patient.

10. The system of claim 2, wherein the instructions further cause the processor to derive one or more timing interval measurements from the synthetic TDI waveform, the one or more714937-9437-8896Docket No. (183161-37)timing interval measurements comprising at least one of: (a) a time from an onset of an R wave on the ECG data to a peak of a systolic velocity wave on the synthetic TDI waveform; (b) a duration of mechanical contraction determined from the onset of the R wave to a peak displacement in systole, wherein the peak displacement is determined as a time integral of the synthetic TDI waveform; (c) a mechanical relaxation time defined as a difference between an RR interval and the duration of mechanical contraction; (d) a time from the onset of the R wave to an onset of a peak early diastolic velocity wave on the synthetic TDI waveform; or (e) a ratio between a QT interval reflecting electrical systole and the time from the onset of the R wave to the peak displacement reflecting mechanical contraction; wherein the one or more timing interval measurements exhibit a statistically significant linear correlation with corresponding timing interval measurements derived from real TDI waveforms acquired from echocardiographic imaging.

11. The system of claim 2, wherein the instructions further cause the processor to: integrate the synthetic TDI waveform over a systolic period of the cardiac cycle to derive a synthetic mitral annular plane systolic excursion (MAPSE) value for the patient.

12. The system of claim 11, wherein the instructions further cause the processor to: estimate a synthetic global longitudinal strain (GLS) value for the patient based at least in part on the synthetic MAPSE value and one or more features derived from the ECG data.

13. A method for generating synthetic echocardiographic waveform data, the method comprising:acquiring cardiac electrical signals detected at one or more surface electrodes positioned about a cardiac region of a patient and representative of cardiac activity of the patient during a given measurement time period;processing the cardiac electrical signals using a first trained machine-learning model, to generate synthetic echocardiographic waveform data;wherein the synthetic echocardiographic waveform data represents, with at least 80% alignment, non-synthetic echocardiographic waveform data obtained via ultrasound imaging from the patient for the cardiac activity during the given measurement period; anddetermining, using a second trained machine learning model, a parameter of at least one cardiac biomechanical function from the synthetic echocardiographic waveform data.724937-9437-8896Docket No. (183161-37)14. The method of claim 13, further comprising generating a digital twin of left ventricular mechanics based on the synthetic echocardiographic waveform data, without use of data obtained via echocardiographic imaging equipment.

15. The method of claim 13, wherein the at least one cardiac biomechanical function comprises cardiac tissue motion dynamics, and the second trained machine learning model was trained to determine myocardial tissue velocity characteristics from synthetic echocardiographic waveform training data.

16. The method of claim 15, wherein the parameter comprises at least one of: tissue Doppler imaging (TDI) velocities, mitral annular plane systolic excursion (MAPSE), or global longitudinal strain (GLS).

17. The method of claim 13, wherein the first trained machine-learning model comprises a generator network trained as a component of a generative adversarial network (GAN) further comprising: (i) a discriminator network trained to distinguish real TDI waveforms from synthetic TDI waveforms generated by the trained generator network; and (ii) an auxiliary regressor trained to predict peak velocity characteristics of TDI waveforms, wherein the discriminator network and the auxiliary regressor each provided training feedback to the trained generator network during training of the GAN.

18. The method of claim 14, wherein the synthetic echocardiographic waveform data comprises tissue Doppler imaging (TDI) velocities including at least one of systolic velocity (s'), early diastolic velocity (e'), or late diastolic velocity (a'), and the second trained machine learning model was trained to determine myocardial tissue velocity characteristics by extracting peak velocity magnitudes and timing measurements from the synthetic echocardiographic waveform data.

19. The method of claim 13, wherein the at least one cardiac biomechanical function comprises left ventricular mechanics, and the second trained machine learning model was trained to assess at least one of left ventricular systolic dysfunction or left ventricular diastolic dysfunction from the synthetic echocardiographic waveform data.

20. The method of claim 19, wherein the left ventricular mechanics further comprise at least one of mitral annular plane systolic excursion (MAPSE) or global longitudinal strain (GLS), and the second trained machine learning model was trained to predict ejection fraction734937-9437-8896Docket No. (183161-37)status based at least in part on the MAPSE or GLS derived from the synthetic echocardiographic waveform data.

21. The method of claim 13, wherein the at least one cardiac biomechanical function comprises myocardial performance, and the second trained machine learning model was trained to stratify mortality risk for the patient based on parameters derived from the synthetic echocardiographic waveform data and to identify at least one cardiac structural or functional abnormality comprising at least one of diastolic dysfunction, aortic stenosis, valvular heart disease, ventricular hypertrophy, or right ventricular systolic dysfunction from the synthetic echocardiographic waveform data without requiring echocardiographic imaging of the patient.

22. The method of claim 13, further comprising determining that one or more surface electrodes do not include a desired number or a desired position of electrodes to generate the parameter, based on a predetermined threshold, and synthesizing additional data from missing electrodes using at least one of: a transformation equation based on known mathematical relationships between at least two of the one or more surface electrodes that are available; or a vectorcardiographic projection.

23. A method for assessing cardiac function from electrocardiogram (ECG) data, the method comprising:acquiring ECG data from one or more surface electrodes in contact with a patient; processing the ECG data via a foundation model to generate a set of representation features of the ECG data;processing the ECG data through a trained generator network of a generative adversarial network (GAN) to generate synthetic cardiac velocity waveform data temporally correlated with cardiac tissue movement of the patient;extracting statistical features from the synthetic cardiac velocity waveform data; deriving at least one cardiac mechanical property of the patient from the representation features and the statistical features; andoutputting a cardiac function assessment based on the at least one cardiac mechanical property.

24. The method of claim 23, wherein deriving the at least one cardiac mechanical property comprises integrating the synthetic cardiac velocity waveform data over a systolic period to derive a synthetic mitral annular plane systolic excursion (MAPSE) value.744937-9437-8896Docket No. (183161-37)25. The method of claim 24, further comprising estimating a synthetic global longitudinal strain (GLS) value based at least in part on the synthetic MAPSE value and one or more features derived from the ECG data.

26. The method of claim 23, wherein deriving the at least one cardiac mechanical property comprises combining predictions from a first machine learning model trained on the statistical features and a second machine learning model trained on the representation features using an ensemble model employing stacking to generate a combined risk score.754937-9437-8896

Images