Artificial intelligence-guided screening of under-recognized cardiomyopathies adapted for point-of-care cardiac ultrasound

Abstract

Provided herein are methods of training a model for cardiac phenotyping using cardiac ultrasonography images and videos adaptable to point-of-care acquisition. The method includes providing an echocardiogram dataset; labeling the echocardiogram dataset with at least one condition of interest; splitting the echocardiogram dataset into a derivation dataset and a testing dataset; initializing a deep neural network (DNN); automating the extraction of echocardiographic view quality metrics; generating natural and synthetic augmentations of cardiac images and videos; and implementing a loss function that accounts for variations in view quality to train noise-adjusted computer vision models for phenotyping at the point-of-care. Also provided herein is a method of cardiac phenotyping employing the model.

Description

[0001] TITLE OF THE INVENTION

[0002] Artificial intelligence-guided screening of under-recognized cardiomyopathies adapted for point-of-care cardiac ultrasound

[0003] CROSS-REFERENCE TO RELATED APPLICATIONS

[0004] The present application claims priority under 35 U.S.C. § 119(e) to U.S. Provisional Patent Application No. 63 / 562,335, filed March 7, 2024, which application is incorporated herein by reference in its entirety.

[0005] BACKGROUND OF THE DISCLOSURE

[0006] Point-of-care ultrasonography (POCUS) enables in-depth cardiac phenotyping at the point-of-care as a direct extension of the physical exam, and is increasingly used as an adjunctive diagnostic tool in outpatient clinics, emergency rooms and inpatient facilities. In recent years, echocardiography has benefited from advances in computer vision and medical artificial intelligence (Al), which not only provide automated summaries of routine transthoracic echocardiograms (TTE), but also augment our ability to phenotype cardiovascular pathology, with improved accuracy relative to standard workflows. However, Al algorithms in echocardiography are almost invariantly developed and validated using videos acquired by expert technicians and interpreted by board-certified readers. Consequently, models trained using these resources may struggle to generalize to point-of-care studies often limited by off-axis views, incomplete protocols, general-use ultrasound systems and less experienced operators.

[0007] With the expanding use of point-of-care ultrasonography and the arrival of new tools that can assist novice operators in acquiring standard echocardiographic views, there is a growing need to adapt Al phenotyping algorithms for efficient screening outside dedicated echocardiographic laboratories. This is particularly important for the screening of underdiagnosed cardiomyopathies that may benefit from timely diagnosis and risk stratification, such as transthyretin amyloid cardiomyopathy (ATTR-CM), aortic stenosis (AS), and hypertrophic cardiomyopathy (HCM). In fact, several studies suggest that only a minority (10- 20%) of these cases are identified clinically, with disparities in outpatient diagnosis and care disproportionately affecting individuals from lower socioeconomic classes and racial / ethnic minorities. Accordingly, there is a need in the art for articles and methods that improve on existing methods by providing efficient echocardiographic Al tools adapted for opportunistic, point-of- care screening as well as prediction of under-diagnosed cardiomyopathies. The present invention addresses this need.

[0008] SUMMARY

[0009] In one aspect, a method of training a model for cardiac phenotyping includes providing an echocardiogram dataset; labeling the echocardiogram dataset with at least one condition of interest; splitting the echocardiogram dataset into a derivation dataset and a testing dataset; initializing a deep neural network (DNN) using random or pre-trained weights; and training a model on the derivation dataset. In some embodiments, the method further includes preprocessing the echocardiogram dataset. In some embodiments, the DNN enables video-level classification by assigning a probability that a given video corresponds to a standard anatomical view.

[0010] In some embodiments, the training step further includes at least one of introducing artificial augmentations, introducing natural augmentations, and customizing the training loss function. In some embodiments, the method further includes defining a quality of the view relative to a standard set of views. In some embodiments, the training step further includes at least one of weighting for class imbalance, label smoothing, synthetic data augmentation, natural data augmentation, and loss function modification. In some embodiments, the weighting for class imbalance comprises computing class-specific weights to account for an imbalance in dataset labels, the weighting for class imbalance including assigning higher weights to rare labels that are under-represented in the cohorts. In some embodiments, the synthetic data augmentation comprises at least one of random zero padding by a given number of pixels in each spatial dimension, random horizontal flipping, and a random rotation. In some embodiments, the synthetic data augmentation accounts for at least one of variable orientation, off-axis, zoomed out, and zoomed in views. In some embodiments, natural data augmentation includes training separate models for each view-of-interest; and training all-inclusive, view-naive models in pooled datasets; wherein the pooled datasets include all standard echocardiographic views with the classifier blinded to the input view (i.e., apical, parasternal, subcostal). In some embodiments, the views-of-interest include key standard cardiac ultrasound views, routinely acquired with portable cardiac ultrasound (parasternal long axis, parasternal short axis, apical views). In some embodiments, the loss function modification further includes weighing the loss function for the label-specific weights and view probability; wherein higher probabilities denote greater anatomical correctness compared to lower ones.

[0011] In some embodiments, the training step forces the model to learn generalizable signatures detectable across low and high-quality studies. In some embodiments, the trained model is applicable to echocardiograms and sonograms. In some embodiments, the at least one condition of interest comprises a cardiac disorder. In some embodiments, the cardiac disorder comprises cardiomyopathies and valve diseases. In some embodiments, the at least one condition of interest comprises transthyretin amyloid cardiomyopathy (ATTR-CM), aortic stenosis (AS), hypertrophic cardiomyopathy (HCM), and combinations thereof. In some embodiments, the trained model predicts a risk of developing the at least one condition.

[0012] In another aspect, a method of cardiac phenotyping includes applying the model trained according to any one of the embodiments disclosed herein to one or more cardiac images. In some embodiments, the one or more cardiac images include cardiac ultrasound images. In some embodiments, the one or more cardiac images are obtained at the point-of-care.

[0013] In another aspect, an apparatus for cardiac phenotyping includes a processor; a memory unit; and a communication interface; wherein the processor is connected to the memory unit and the communication interface; and wherein the processor and memory are configured to implement the method of cardiac phenotyping according to any of the embodiments disclosed herein.

[0014] In another aspect, provided herein is a computer readable storage medium storing computer-executable instructions for performing the method of cardiac phenotyping according to any of the embodiments disclosed herein.

[0015] BRIEF DESCRIPTION OF THE DRAWINGS

[0016] FIGS. 1A-B show images illustrating study overview. (A) Study hypothesis. (B) Overview of study design and datasets. Al: artificial intelligence; AS: (severe) aortic stenosis; ATTR-CM: amyloid transthyretin cardiomyopathy; ED: emergency department; HCM: hypertrophic cardiomyopathy; POCUS: point-of-care ultrasound; TTE: (standard) transthoracic echocardi ography . FIGS. 2A-B show graphs illustrating video and study-level performance of a viewagnostic multi-label deep learning algorithm. (A) Video-level performance (across all available parasternal long, parasternal short and all apical views) for discrimination of HCM, ATTR-CM and AS. (B) Study-level performance by simple mean averaging of all available videos within a given study. AS: (severe) aortic stenosis; ATTR-CM: amyloid transthyretin cardiomyopathy; AUROC: area under the receiver operating characteristic curve; CI: confidence interval; HCM: hypertrophic cardiomyopathy.

[0017] FIGS. 3A-D show graphs illustrating video and study-level performance of a viewagnostic multi-label deep learning algorithm applied to POCUS. (A) Video-level performance (AUROC with 95% CI) for discrimination of HCM, ATTR-CM, and AS, by deploying a POCUS-adapted, view-agnostic model to different echocardiographic views obtained in the ED (blue = PLAX; orange = PSAX at the papillary muscle level; green = A4C; and red = any of the 3 views combined). (B) Results presented for videos with an automated view classifier certainty of 50% of greater. (C-D) Study-level results obtained by simple mean averaging of the videolevel predictions for each study. A4C: apical-4-chamber view; AS: (severe) aortic stenosis; ATTR-CM: amyloid transthyretin cardiomyopathy; AUROC: area under the receiver operating characteristic curve; CI: confidence interval; HCM: hypertrophic cardiomyopathy; PLAX: parasternal long axis view; POCUS: point-of-care ultrasonography; PSAX: parasternal short axis view.

[0018] FIGS. 4A-F show images illustrating saliency maps. Activation maps for HCM (A-B), ATTR-CM (C-D) and AS (E-F) across PLAX and A4C views obtained at the point-of-care in the emergency department. A4C: apical -4-chamber view; AS: (severe) aortic stenosis; ATTR- CM: amyloid transthyretin cardiomyopathy; HCM: hypertrophic cardiomyopathy; PLAX: parasternal long axis view.

[0019] FIGS. 5A-F show graphs illustrating video and study-level performance of a viewagnostic multi-label deep learning algorithm across distinct views. (A-C) Video-level performance (PLAX (A), PSAX (B), and A4C (C) views) for discrimination of HCM, ATTR- CM and AS. (D-F) Study-level performance (PLAX (D), PSAX (E), and A4C (F) views) by simple mean averaging of all available videos within a given study and for each individual view. A4C: apical-4-chamber view; AS: (severe) aortic stenosis; ATTR-CM: amyloid transthyretin cardiomyopathy; AUROC: area under the receiver operating characteristic curve; CI: confidence interval; HCM: hypertrophic cardiomyopathy; PLAX: parasternal long axis view; PSAX: parasternal short axis view.

[0020] FIGS. 6A-B show graphs illustrating head-to-head video-level and study-level performance of a view-agnostic vs view-specific multi-label deep learning algorithm. (A) Videolevel performance (across all available PLAX, PSAX (papillary muscle level) and A4C views) for discrimination of HCM (blue), ATTR-CM (red) and AS (green) using view-agnostic (uninterrupted line) vs view-specific models (dotted lines). (B) Study-level performance by simple mean averaging of all available videos within a given study. A4C: apical-4-chamber view; AS: (severe) aortic stenosis; ATTR-CM: amyloid transthyretin cardiomyopathy; AUROC: area under the receiver operating characteristic curve; CI: confidence interval; HCM: hypertrophic cardiomyopathy; PLAX: parasternal long axis view; PSAX: parasternal short axis view.

[0021] FIG. 7 shows a graph illustrating a density plot of view classifier probabilities. Density plot of view classifier probabilities across the TTE testing set (purple) and the POCUS cohort, presented separately for HCM, ATTR-CM, AS, and control cases. Higher probabilities denote greater certainty in matching a given video to a standardized echocardiographic view. AS: (severe) aortic stenosis; ATTR-CM: amyloid transthyretin cardiomyopathy; HCM: hypertrophic cardiomyopathy; TTE: transthoracic echocardiography.

[0022] FIGS. 8A-L show images illustrating true versus false positive and negative predictions across labels and views in the ED POCUS cohort. Representative frames of the top 5 true positive, true negative, false positive and false negative cases for each label of interest (A-D: HCM; E-H: TTR; I-L: AS) and view. A4C: apical -4-chamber view; AS: (severe) aortic stenosis; ATTR-CM: amyloid transthyretin cardiomyopathy; HCM: hypertrophic cardiomyopathy; PLAX: parasternal long axis view; PSAX: parasternal short axis view.

[0023] DETAILED DESCRIPTION OF THE DISCLOSURE

[0024] Definitions

[0025] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Although any methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present invention, the preferred methods and materials are described.

[0026] The articles “a” and “an” are used herein to refer to one or to more than one (z.e., to at least one) of the grammatical object of the article. By way of example, “an element” means one element or more than one element.

[0027] “About” as used herein when referring to a measurable value such as an amount, a temporal duration, and the like, is meant to encompass variations of ±20% or ±10%, more preferably ±5%, even more preferably ±1%, and still more preferably ±0.1% from the specified value, as such variations are appropriate to perform the disclosed methods.

[0028] Ranges: throughout this disclosure, various aspects of the invention can be presented in a range format. It should be understood that the description in range format is merely for convenience and brevity and should not be construed as an inflexible limitation on the scope of the invention. Accordingly, the description of a range should be considered to have specifically disclosed all the possible subranges as well as individual numerical values within that range. For example, description of a range such as from 1 to 6 should be considered to have specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6 etc., as well as individual numbers within that range, for example, 1, 2, 2.7, 3, 4, 5, 5.3, and 6. This applies regardless of the breadth of the range.

[0029] Detailed Description

[0030] Provided herein are methods of training a model for cardiac phenotyping. In some embodiments, the method includes providing an echocardiogram dataset; labeling the echocardiogram dataset with at least one condition of interest; splitting the echocardiogram dataset into a derivation dataset and a testing dataset; initializing a deep neural network (DNN) using random or pre-trained weights; and training a model on the derivation dataset. In some embodiments, the method further includes automating the extraction of echocardiographic view quality metrics; generating natural and synthetic augmentations of cardiac images and videos; and / or implementing a loss function that accounts for variations in view quality to train noise- adjusted computer vision models for phenotyping at the point-of-care. In some embodiments, the derivation dataset comprises a training dataset and a validation dataset.

[0031] In some embodiments, the method includes pre-processing one or more sub-sets (e.g, derivation dataset, testing dataset, and / or any other sub-set) of the echocardiogram dataset. In some embodiments, the pre-processing includes loading pixel data and one or more of masking pixels in a periphery to remove identifying information; converting to an appropriate image or video-based format; sampling one or more frames (e.g., between 1 and 100 frames) from each video either at random or in a pre-defined sequence (e.g., every second, third, .. ,nthframe); downsampling the frames to a lower resolution e.g., 224x224 pixels), forming downsampled frames; and / or feeding the frames through a view classifier rating the anatomical correctness relative to standard echocardiographic views. In some embodiments, the pre-processing further includes at least one of binarizing each video frame with a fixed threshold, masking out all pixels outside a convex hull of a largest contour, and / or downsampling to 112x112 pixels as .AVI files.

[0032] In some embodiments, the training step includes at least one of introducing artificial augmentations, introducing natural augmentations, and / or customizing the training loss function. For example, in some embodiments, the training step includes at least one of weighting for class imbalance, label smoothing, synthetic data augmentation, natural data augmentation, and loss function modification. In some embodiments, the training step includes defining a quality of a view relative to a standard set of views.

[0033] In some embodiments, weighting for class imbalance comprises computing class-specific weights to account for an imbalance in dataset labels, the weighting for class imbalance including assigning higher weights to rare labels that are under-represented in the cohorts (e.g., ATTR-CM).

[0034] In some embodiments, the synthetic data augmentation comprises at least one of random zero padding by a given number of pixels in each spatial dimension, random horizontal flipping, and a random rotation. A suitable number of pixels includes, but is not limited to, up to 50 pixels, more than 50 pixels, up to 30 pixels, between 8 and 16 pixels, or any combination, subcombination, range, or sub-range thereof. A suitable rotation for the random rotation includes, but is not limited to, between -90 and +90 degrees, between -45 and +45 degrees, between -30 and +30 degrees, between -20 and +20 degrees, between -15 and +15 degrees, between -10 and +10 degrees, or any combination, sub-combination, range, or sub-range thereof. In some embodiments, the synthetic data augmentation accounts for at least one of variable orientation, off-axis, zoomed out, and zoomed in views.

[0035] In some embodiments, the natural data augmentation comprises training separate models for each view-of-interest, and training all-inclusive, view-naive models in pooled datasets. The pooled datasets include any suitable datasets, such as, but not limited to, all standard echocardiographic views, with the classifier blinded to the input view (e.g., apical, parasternal, subcostal). In some embodiments, the views-of-interest include key standard echocardiographic views, such as those routinely acquired with portable echocardiography (e.g., parasternal long axis, parasternal short axis, apical views). Additionally, in embodiments where the quality of a view is defined relative to a standard set of views, model development can be weighed with the view quality as well.

[0036] In some embodiments, the loss function modification comprises a modified version of binary cross entropy (BCE), with or without logits loss function. In some embodiments, the modified version of BCE combines a sigmoid layer and the BCE loss function in one class, making it more numerically stable by taking advantage of the log-sum-exp trick for numerical stability. In some embodiments, the loss function modification includes weighing the loss function for the label-specific weights and view probability, where higher probabilities (e.g., PLAX view probability of 1.00) denote greater anatomical correctness compared to lower ones. The view probability includes the highest numerical probability of the view classifier that defines the most likely view. In some embodiments, the loss function modification penalizes the model for missing under-represented labels, particularly in the context of a challenging view. Additionally or alternatively, in some embodiments, customizing the training loss function boosts the model to identify key phenotypes from the hardest cases, thus learning to recognize off-axis and limited view images, and identify conditions that human readers cannot perceive on these 2-dimensional ultrasound videos of the heart.

[0037] The training step may include any suitable parameters based upon the data sets being used. For example, in some embodiments, the training step includes one or more of a maximum of a defined number of epochs with early stopping such that if the average validation performance metric for the labels of interest does not improve after a given number of consecutive epochs; a fixed or adaptive learning rate of 10'6to 10’4; and / or using sampled videos clips of a given number of frames either at random or by identifying every second, third, fourth, fifth, . . ., nthframe. Any suitable number of epochs may be used, such as, but not limited to, between 10 and 300 epochs. The number of consecutive epochs may include, but is not limited to, between 3 and 5. In some embodiments, the training step further includes applying padding with empty frames along the temporal axis if either the video was too short or the randomly chosen starting point of the clip was near the end of the video. In some embodiments, the DNN described herein enables video-level classification by assigning a probability that a given video corresponds to a standard anatomical view. In some embodiments, the probabilities add up to 1 across all views. In some embodiments, the training step forces the model to learn generalizable signatures detectable across low and high-quality studies. Additionally or alternatively, the trained model is applicable to any suitable type of data and / or view, such as, but not limited to, echocardiograms and sonograms.

[0038] The method disclosed herein may be applied to and / or for phenotyping of any suitable condition of interest detectable in the echocardiogram dataset. For example, in some embodiments, the at least one condition of interest comprises a cardiac disorder, including structural and functional anomalies of the heart. In some embodiments, the cardiac disorder comprises cardiomyopathies and valve diseases. For example, suitable cardiac disorders include, but are not limited to, transthyretin amyloid cardiomyopathy (ATTR-CM), aortic stenosis (AS), hypertrophic cardiomyopathy (HCM), and / or combinations thereof.

[0039] After training, the trained model can be applied to phenotype any of the one or more conditions of interest. Additionally or alternatively, the trained model can be applied to predict a risk of developing any of the one or more conditions of interest. Accordingly, also provided herein are methods of cardiac phenotyping, the methods comprising applying the model trained according to any of the embodiments disclosed herein. The model can be applied to any suitable cardiac image, such as, but not limited to, echocardiograms, sonograms, and combinations thereof. In some embodiments, the cardiac image(s) are obtained at the point-of-care. In some embodiments, the method includes detecting a condition of interest that otherwise requires advanced diagnostics not typically possible due to technical and skill limitations on hand-held devices. Additionally or alternatively, in some embodiments, the method includes applying the model to real-world portable videos (e.g., ultrasound), and detecting cardiac conditions that are often invisible to the untrained eye. Accordingly, in some embodiments, the methods disclosed herein enable the timely screening of common and rare medical conditions, such as rare cardiomyopathies, directly in outpatient clinics or the emergency room. In such embodiments, patients can then be referred for confirmatory testing to prevent subsequent morbidity and mortality.

[0040] Also provided herein is an apparatus for cardiac phenotyping, the apparatus comprising a processor, a memory unit, and a communication interface. The processor is connected to the memory unit and the communication interface, and the processor and memory are configured to implement the method according to any of the embodiments disclosed herein.

[0041] Also provided herein is a computer-readable storage medium storing computerexecutable instructions for performing any of the methods disclosed herein.

[0042] Those skilled in the art will recognize or be able to ascertain using no more than routine experimentation, numerous equivalents to the specific procedures, embodiments, claims, and examples described herein. Such equivalents are considered to be within the scope of this invention and covered by the claims appended hereto.

[0043] It is to be understood that wherever values and ranges are provided herein, all values and ranges encompassed by these values and ranges, are meant to be encompassed within the scope of the present invention. Moreover, all values that fall within these ranges, as well as the upper or lower limits of a range of values, are also contemplated by the present application.

[0044] The following examples further illustrate aspects of the present invention. However, they are in no way a limitation of the teachings or disclosure of the present invention as set forth herein.

[0045] EXAMPLES

[0046] EXAMPLE 1

[0047] Point-of-care ultrasonography (POCUS) enables in-depth cardiac phenotyping at the point-of-care as a direct extension of the physical exam, and is increasingly used as an adjunctive diagnostic tool in outpatient clinics, emergency rooms and inpatient facilities. In the recent years, echocardiography has benefited from advances in computer vision and medical artificial intelligence (Al) which not only provide automated summaries of routine transthoracic echocardiograms (TTE), but also augment our ability to phenotype cardiovascular pathology, with improved accuracy relative to standard workflows. However, Al algorithms in echocardiography are almost invariantly developed and validated using videos acquired by expert technicians and interpreted by board-certified readers. Consequently, models trained using these resources may struggle to generalize to point-of-care studies that are often limited by off- axis views, incomplete protocols, general use ultrasound systems and less experienced operators.

[0048] With the expanding use of point-of-care ultrasonography and the arrival of new tools that can assist novice operators in acquiring standard echocardiographic views, there is a growing need to adapt AT phenotyping algorithms for efficient screening outside dedicated echocardiographic laboratories. This is particularly important for the screening of underdiagnosed cardiomyopathies that may benefit from timely diagnosis and risk stratification, such as transthyretin amyloid cardiomyopathy (ATTR-CM), aortic stenosis (AS), and hypertrophic cardiomyopathy (HCM). In fact, several studies suggest that only a minority (10- 20%) of these cases are identified clinically, with disparities in outpatient diagnosis and care disproportionately affecting individuals from lower socioeconomic classes and racial / ethnic minorities.

[0049] This Example describes the development and testing of a framework for the efficient training of echocardiographic Al tools specifically adapted for opportunistic, point-of-care screening of under-diagnosed cardiomyopathies (FIG. 1A). Using a large library of transthoracic echocardiograms, we implement a customizable training pipeline involving a series of synthetic and natural augmentation steps along with customized weights reflecting the anatomical correctness of the input videos, thus allowing our models to learn generalizable phenotypic signatures of each condition from widely available standard, two-dimensional echocardiograms. We go on to demonstrate the generalizability of this approach to real-world POCUS videos obtained as part of clinical evaluation in the Emergency Departments of a large hospital network, highlighting the untapped potential of opportunistic screening at the point-of-care.

[0050] Methods

[0051] Study Population and Data Source

[0052] The study was designed as a case-control analysis of patients seeking care across five distinct hospitals affiliated with the Yale-New Haven Health (YNHH) system across Connecticut and Rhode Island.

[0053] TTE (development) cohort (Yale -New Haven Health [YNHH] echo laboratories)

[0054] We queried the electronic health record (EHR) and linked echocardiographic library (n=522,507 unique studies) of YNHH for transthoracic echocardiography (TTE) exams performed between 5 / 12 / 2015 (database inception) and 12 / 31 / 2022 in patients with HCM, ATTR-CM and AS (FIG. IB). We defined three, non-mutually exclusive labels of interest, as follows: i) HCM: We screened for all individuals with an ICD-9 / ICD-10 code for any cardiomyopathy (425, 142.0, 142.1, 142.2, 142.5, 142.8, 142.9, 143.1, 143.8) or heart failure (428, 150*), inclusive of HCM-specific ICD-9 / 10 codes (425.1, 425.11, 425.18, 142.1, 142.2); to maximize the specificity of our definition given the known unreliability of billing / administrative codes in accurately capturing patient phenotypes, we required positive labels for this condition to have undergone cardiac magnetic resonance imaging with the final interpretation / conclusion supporting the presence of the diagnosis. Given that HCM is a genetic cardiomyopathy, we included all available echocardiograms regardless of their timing relative to the time of diagnosis. ii) Amyloid cardiomyopathy (ATTR-CM): We screened for all individuals with an ICD- based diagnosis of any cardiomyopathy or heart failure as above or an amyloidosisspecific code (277.3, 277.30, 277.39, E85.2, E85.82, E85.4, E85.8, E85.9, excluding E85.81 [light chain amyloidosis]). Similar to HCM, to increase the specificity of the label we required positive labels to have undergone bone scintigraphy (with 99mlc- pyrophosphate) which was interpreted as positive for cardiac uptake by the interpreting physician (i.e., semi -quantitative visual score of 2 or 3 or heart to contralateral lung ratio >1.5). For positive cases we defined the time of diagnosis as the time of the positive PYP scan, and, to account for the delay between disease onset and diagnosis (median delay of ~13 months as previously reported in the literature), we included echocardiograms performed up to 12 months before this date (and any time after).

[0055] Hi) Severe AS: Similar to our prior work, we identified cases of severe AS, inclusive of severe low-flow, low-gradient AS, based on the interpretation of a TTE exam by a board- certified reader and in agreement with existing guidelines.

[0056] Controls were extracted by randomly sampling echocardiograms from the same period, further enriched with consecutive individuals who underwent TTE in 2021, extracted as previously described (Holste, G. et al. Severe aortic stenosis detection by deep learning applied to echocardiography. Eur. Heart J. (2023) doi:10.1093 / eurheartj / ehad456; Oikonomou, E. K. et al. A digital biomarker for aortic stenosis development and progression using deep learning for two-dimensional echocardiography. medRxiv (2023) doi:10.1101 / 2023.09.28.23296234.). To qualify for inclusion in the control arm, we required the absence of any HCM-specific ICD codes (425.1, 425.11, 425.18, 142.1, 142.2) or MRI findings confirming HCM, any amyloidosis-related ICD codes and no evidence of severe AS on corresponding echocardiography.

[0057] After following the above process, we identified a cohort of 28,387 unique TTE studies in 16,487 unique patients with a total of 874,972 unique videos.

[0058] Point-of-care cardiac ultrasonography (POCUS) cohort (YNHH Emergency Departments) :

[0059] After defining participant- and patient-specific labels, as described above, we queried the picture archiving and communication system (PACS) of the hospital system for any of these 16,487 participants who underwent a point-of-care emergency cardiac ultrasound by an emergency room provider. Such studies are performed by emergency room staff using a general- purpose ultrasound system (Sparq Ultrasound system, Philips) and are not read by echocardiographers. We excluded any POCUS studies performed after the onset of end-stage renal disease (‘585.6’, ‘N18.6’), after heart transplantation (‘V42.1’, ‘Z94.1 ’, ‘Z94.3), and / or aortic valve replacement (ICD procedure codes ‘35.21’, ‘35.22’, ‘02RF07Z’, ‘02RF08Z’, ‘02RF0JZ’, ‘02RF0KZ’), as well as studies that did not contain any of the following echocardiographic views (parasternal long axis [PLAX], parasternal short axis at the papillary muscle level [PSAX] and / or apical four-chamber [A4C]), our screening yielded 1,879 unique studies in 1,330 unique patients with 3,758 PLAX / PSAX / A4C performed between 2 / 23 / 2011 and 10 / 30 / 2023.

[0060] Defining training, validation and testing splits

[0061] To avoid data leakage, we removed all TTE studies of all patients with any available POCUS scans from our TTE cohort. After assigning HCM, ATTR-CM, and AS labels as reviewed above, we randomly split our TTE cohort into a training 75%, validation (15%) and internal testing (10%) set. This ensured that distinct echocardiograms belonging to the same patient would not be present across more than one of the three (training, validation, testing) sets. To further boost the validity of the observations, we excluded any echocardiographic study with a measured interventricular septal thickness during diastole (IVSd) of 1.3 cm or greater from our training set. However, this was not done during validation or testing to ensure a reliable assessment of the model’s performance.

[0062] Automated view pre-processing and view classification

[0063] All studies (TTE and POCUS) were first exported as .DICOM files. We implemented an end-to-end pre-processing pipeline that we have previously published (Holste, G. et al. Severe aortic stenosis detection by deep learning applied to echocardiography. Eur. Heart J. (2023) doi: 10.1093 / eurheartj / ehad456; Holste, G., Oikonomou, E. K., Mortazavi, B., Wang, Z. & Khera, R. Self-supervised learning of echocardiogram videos enables data-efficient clinical diagnosis. arXiv [cs.CV] (2022).), that involves loading the pixel data, masking our pixels in the periphery to remove identifying information and converting to Audio Video Interleave ( AVI) format. We then randomly sample ten frames from each video, downsample to 224x224 pixels, and feed through a VGG19 convolutional neural network (CNN) that enables video-level classification of 18 echocardiographic views by assigning a probability that a given video corresponds to a standard anatomical view (with probabilities adding up to 1 across all views) . Next, we perform more thorough cleaning and de-identification by binarizing each video frame with a fixed threshold, mask out all pixels outside the convex hull of the largest contour, and downsample to 112x112 pixels as .AVI files for fast loading during training and inference.

[0064] Model Training

[0065] We designed a training framework that incorporated a naive approach to using multiple views (apical, parasternal long, parasternal short, and subcostal views) without annotations as well as a customized training loss to heavily favor low-quality, off-axis videos of patients with the labels of interest. We first initialized a 3D-ResNetl8 CNN architecture by using pre-trained weights from the Kinetics-400 dataset and further modified the output layer of the label to enable multi-label classification for the three representative labels of interest, namely HCM, ATTR-CM, and AS. With the goal of deploying our model in POCUS scans for screening of relatively rare cardiomyopathies, we implemented a range of customizations:

[0066] Natural and synthetic data augmentation methods'. We trained both separate models for each key views-of-interest, namely PLAX, PSAX and A4C, followed by all-inclusive, view- naive models trained in pooled datasets that included all parasternal (long and short) and apical views with the classifier blinded to the input view. This enabled a head-to-head comparison of how view-specific versus view-agnostic approaches generalize to real-world POCUS acquisitions. We further applied a series of data augmentations to account for variable orientation and off-axis views that included random zero padding by up to 8 pixels in each spatial dimension, random horizontal flipping with (probability 0.5), and a random rotation within -10 and 10 degrees (probability 0.5). After augmentation, each video clip’s intensities were normalized to 0-1 and standardized using the channel -wise means and standard deviations from the Kinetics-400 training dataset.

[0067] Quality-adjusted weights and loss function: We defined a loss function to prompt the model to learn from lower quality cases. We took the binary cross entropy (BCE) with logits loss function, that combines a sigmoid layer and the BCE loss function in one class, and incorporated both label-specific weights to account for rare labels, as well as inverse weighting based on the view alignment probabilities. Higher probabilities (i.e., PLAX view probability of 1.00) denote greater anatomical correctness compared to lower ones. The aim was to penalize the model for missing under-represented labels, especially in the context of a challenging view. We also applied label smoothing (<z=0.1) to penalize over-confidence in our model’s predictions.

[0068] Modified loss function description:

[0069] Step 1: View Weights:

[0070] For each sample z in the batch:

[0071] 1 Vi =(Pj + io-5)2

[0072] Where Ptis the view probability for sample i. These weights are then normalized across the batch:

[0073] Where N is the number of samples in the batch.

[0074] Step 2: Class Weights:

[0075] For each label £ in a sample i:

[0076] Cik = T'ikxlk+ (1 — Tik) x Wok

[0077] Where Tikis the true binary label for label k in sample i, and I / F1 / cand Wokare the class weights for the positive and negative classes of label k, respectively. Step 3: Combined Weights:

[0078] The combined weight for each label in each sample is the product of the view weight and the class weight, normalized across the batch:

[0079] Where K is the number of labels.

[0080] Step 4: Weighted Loss:

[0081] Finally, the weighted loss for the batch is the sum of the individual weighted losses:

[0082] Where Likis the binary cross-entropy loss for label k in sample i.

[0083] All models were trained for a maximum of 30 epochs with early stopping such that if the average validation AUROC for the three labels of interest did not improve for 5 consecutive epochs training was terminated and the weights from the epoch with maximum validation AUROC were used for final evaluation. Models were trained on four NVIDIA Tesla T4 GPU with the Adam optimizer, a learning rate of IO"4, a batch size of 56 to maximize GPU utilization, using randomly sampled videos clips of 16 frames and sampling one out of every five frames to enable a global capture of the cardiac cycle (median number of frames 61 25th-75thpercentile: SO- 85]). We applied optional padding with empty frames along the temporal axis if either the video was too short or the randomly chosen starting point of the clip was near the end of the video.

[0084] Model performance assessment

[0085] At the time of inference, we averaged clip-level predictions to obtain video-level predictions for each label. We analyzed the performance of both video-level and study-level predictions, by averaging class-specific probabilities from all videos acquired during the same study. We evaluated the discriminatory performance for HCM, ATTR-CM and AS using standard metrics, including AUROC, the area under the precision-recall curve (AUPRC), as well as metrics across the thresholds that maximized Youden’s J (sum of sensitivity and specificity - 1) as well as at 90% sensitivity, including the Fl score, positive (PPV) and negative predictive value (NPV), and diagnostic odds ratio at 3% prevalence, an estimated average of the prevalence of HCM, ATTR-CM, AS in non-randomly selected individuals with a mean age of >65 years with known or suspected cardiovascular disease.

[0086] Model explainability

[0087] To assist with explainability, we generated sample saliency maps for the most confident HCM, ATTR-CM and AS cases using Gradient-weighted Class Activation Mapping (Grad- CAM) (Selvaraju, R. R. et al. Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization. arXiv [cs.CV] (2016).). We present the pixelwise maximum along the temporal axis to capture the most salient regions.

[0088] Statistical analysis

[0089] Categorical variables are summarized as counts (valid percentages), and continuous variables as mean ± standard deviation, or median [25th-75thpercentile], unless specified otherwise. Categorical variables were compared across distinct groups using the %2test. Continuous variables were compared using the non-parametric Mann-Whitney or Kruskal-Wallis tests for two or three or more groups, respectively. Metrics of discrimination are accompanied by 95% confidence intervals derived from bootstrapping with r=200 replications. Comparisons of AUROC between competing models within the same dataset and for the same label are performed by DeLong’s test. The probability distribution of continuous variables across distinct groups is visualized using density plots, with values compared by the non-parametric Mann- Whitney test. Analyses were performed using Python 3.9.7, using pytorch 1.8.0, torchvision 0.9.0, scipy 1.7.3. All statistical tests were two-sided with a significance level of 0.05, unless specified otherwise.

[0090] Results

[0091] Study population

[0092] The TTE cohort included a total of 10,702 studies with 290,245 echocardiographic videos among 8,460 unique patients, split into a training (n=8,090), validation (n=l,577) and a testing set (1,035). In the training set, the mean age at the time of echocardiography was 68.3±15 years, 3,752 (46.4%) were women, 6,087 (75.2%) reported White race, 612 (7.6%) Black race and (433) 5.4% Hispanic / Latinx ethnicity. In total, 1,375 studies corresponded to HCM (17.0%), 228 (2.8%) to ATTR-CM and 1, 120 (15.0%) to severe AS, with comparable prevalence among the training, validation, and testing splits (Table 1). The ED POCUS population included 1,879 studies with 3,758 PLAX, PSAX, or A4C views (mean age at the time of the ED visit was 68.3±16.7 years). Likely reflecting differential access to emergent / urgent versus routine outpatient care in the community, the prevalence of female sex, and non-white race or Hispanic ethnicity was significantly higher in the ED POCUS than the TTE cohort; in summary, 1,053 studies (56.0%) performed in women, 363 (19.3%) in Black and 198 (10.5%) in Hispanic individuals (p<0.001 for all relative to the TTE cohort). There were 83 (4.4%) scans in patients with HCM, 56 with ATTR-CM (16.1%) and 302 (16.1%) with AS.

[0093] Table 1 | Summary of cohort characteristics. Transthoracic echocardiography - ED POCUS TTF Training Validation . testing testing

[0094] Available videos and studies

[0095] Number of unique studies 8,090 1,577 1,035 1,879

[0096] Number of unique videos 218,725 43,031 28,489 3,758

[0097] Parasternal long-axis view (PLAN) 25,569 4,987 3,283 836

[0098] PSAX (papillary muscle level) 26,665 5,114 3,313 2,442

[0099] Apical 4-chamber (A4C) 21,328 4,200 2,649 480

[0100] Number of unique patients 6,376 1,245 839 1,330

[0101] Demographic characteristics (study-level)

[0102] Age, mean (SD) 68.3 (15.0) 68.7 (14.8) 69.2 (15.3) 70.3 (16.7)

[0103] Female gender (%), n (%) 3,752 (46.4) 709 (45.0) 496 (47.9) 1,053 (56.0)

[0104] Hispanic Ethnicity, n (%) 433 (5.4) 80 (5.1) 51 (4.9) 198 (10.5)

[0105] Asian Race, n (%) 109 (1.3) 23 (1.5) 12 (1.2) 20 (1.1)

[0106] Black or African American Race, n (%) 612 (7.6) 98 (6.2) 73 (7.1) 363 (19.3)

[0107] White or Caucasian Race, n (%) 6,087 (75.2) 1,231 (78.1) 813 (78.6) 1,247 (66.4)

[0108] Labels of interest (study-level) Severe Aortic Stenosis, n (%) 1,120 (15.0) 238 (16.0) 152 (15.5) 302 (16.1)

[0109] Hypertrophic Cardiomyopathy (HCM), n (%) 1,375 (17.0) 243 (15.4) 168 (16.2) 83 (4.4)

[0110] Cardiac Amyloidosis, n (%) 228 (2.8) 39 (2.5) 31 (3.0) 56 (3.0)

[0111] Variables summarized as mean (standard deviation [SD]) or counts (percentages), as appropriate.

[0112] Performance of multi-view models for detection of HCM, ATTR-CM, AS on TTE Overall performance: When applied in the internal TTE testing set and averaged across all acquired video clips in a given study, a multi-label, view-agnostic algorithm trained using all parasternal long, parasternal short and apical views was able to reliably discriminate HCM (AUROC 0.95 [95% CI: 0.94-0.96]), ATTR-CM (0.98 [95% CI: 0.96-0.99]) and severe AS (AUROC 0.94 [95% CI: 0.93-0.95]) (FIGS. 2A-B).

[0113] Video-level performance: In a per-video analysis, there was variability in performance with the highest AUROC for HCM and ATTR-CM seen in A4C videos (AUROC 0.90 [95% CI 0.89-0.92] and 0.98 [95% CI 0.97-0.99]), respectively), whereas for AS the highest performance was observed in the PLAX view (AUROC 0.92 [95% CI: 0.91-0.93], when averaging all views in a study) (FIGS. 5A-F).

[0114] View-agnostic vs view-specific models: In an analysis of commonly obtained POCUS views in the TTE cohort (PLAX, PSAX and A4C), the view-agnostic model performed consistently better than an approach that relied on predictions from three separate models trained for each individual view. More specifically, in a per-video analysis, the AUROC for the viewagnostic vs view-specific models for the 3 principal views were as follows: HCM (0.86 [95%CI: 0.85-0.87] vs 0.81 [95%CI: 0.80-0.82]); ATTR-CM: 0.95 [95%CI: 0.94-0.96] vs 0.92 [95%CI: 0.90-0.93]); and AS (0.84 [95%CI: 0.83-0.85] vs 0.79 [95%CI: 0.78-0.80], p<0.001) (FIGS. 6A- B)

[0115] Cardiac-focused POCUS assessment in the emergency room

[0116] As opposed to the TTE cohort where the ratio of PLAX:PSAX:A4C views was almost 1 : 1 : 1 (reflecting a standardized protocol) (Table 1), we observed that more than two in every three (2,442 out of 3,758 [70.0%]) eligible cardiac POCUS videos included the PSAX view, followed by the PLAX (836 [22.2%]) and A4C views (480 [12.8%]). Furthermore, the automated view classifier reported significant lower probabilities (median 0.63 [IQR: 0.44-0.88] vs 0.93 [IQR: 0.69-1.00] in the TTE cohort, pMann-whitney<Q .001), reflecting lower certainty in matching the obtained image to a standardized view. Of note, within the POCUS cohort, there was no significant difference in the view classifier probabilities among videos obtained from studies in patients with HCM, ATTR-CM, AS and controls (pKmskai-waiiis=0.07) (FIG. 7).

[0117] Performance of multi-view, multi-label models for detection of HCM, ATTR-CM, AS using point-of-care echocardiography

[0118] Similar to our observations in the TTE testing set, the view-agnostic model was associated with notable gains in discrimination relative to a view-specific model when deployed in the POCUS cohort, with 5[AUR0C] of 0.06 ( ?=0.08), 0.09 (p=0.009) and 0.06 (p=0.003) for HCM, ATTR-CM and AS, respectively (FIGS. 8A-L).

[0119] The performance of the view-agnostic classifier in the POCUS cohort varied across echocardiographic views as well as input video quality as assessed by the view classifier probability, with gains in performance for all 3 labels when restricting to view classifier probabilities of at least 0.5 (n=2496 from 3,758 videos) (FIGS. 3A-D). For instance, we demonstrate that at a single video level, our classifier identified ATTR-CM and severe AS from PLAX views with an AUROC of 0.94 and 0.81, respectively, and HCM from A4C views with an AUROC of 0.85. At the threshold level that maximized Youden’s J (sum of sensitivity and specificity - 1) these thresholds corresponded to a diagnostic odds ratio of 46.7 (for ATTR-CM), 9.2 (for severe AS) and 25.5 for HCM, respectively (see Table 2 for threshold-specific summaries of sensitivity, specificity and PPV / NPV metrics).

[0120] Table 2 | POCUS-based discrimination of HCM, ATTR-CM and AS.

[0121] A4C: apical 4-chamber; AS: aortic stenosis: ATTR-CM: transthyretinamyloid cardiomyopathy; HCM: hypertrophic cardiomyopathy; OR: odds ratio; PLAN: parasternal long axis; POCUS: point-of-care ultrasonography: PS AX: parasternal short axis (papillary muscle level). * PPV and NPV reported at simulated 3% prevalence.

[0122] Explainability and representative cases

[0123] Grad-CAM: Representative Grad-CAM maps for the top (true positive predictions) for each label are shown in FIGS. 4A-F. For HCM these generally localized to the left ventricle, for ATTR-CM, the signal seems to localize to the left atrium, whereas for severe AS the focuses seems to be on the left ventricle and the aortic valve (when in plane), consistent with our prior work.

[0124] Representative cases: In FIGS 8A-L we provide an illustrative summary of studies corresponding to the top five highest and lowest predictions for both cases and controls across all three key views (PLAX, PSAX, A4C). These examples showcase the key challenges of POCUS imaging, including the variation in acquisition protocols, probe orientation, off-axis views, and significant noise artifacts.

[0125] Discussion

[0126] In the present study we demonstrate that Al algorithms trained using large datasets of standard transthoracic echocardiograms (TTE) can be adapted for use with point-of-care ultrasonography (POCUS), thus enabling opportunistic screening of rare cardiomyopathies across a range of settings with images acquired by non-expert operators. Our approach describes a preprocessing framework and training pipeline that exposes video-based models to the unique challenges of handheld echocardiography, such as off-axis views and variable acquisition protocols, thus forcing the models to learn generalizable signatures detectable across low and high-quality studies. For instance, we show that view-agnostic models, models that are exposed to the whole range of views that are available during training, consistently outperform viewspecific models, a pattern that persists across both standard TTE and POCUS videos. The feasibility and superior performance of this approach is illustrated across more than 3,000 POCUS videos obtained in the emergency rooms of a hospital health system over a 13-year period, with POCUS-adaptable convolutional neural networks effectively discriminating cases of HCM, ATTR-CM and AS on single echocardiographic videos. Taken together, these findings provide a guide on how POCUS-adaptable training of Al echocardiographic algorithms can maximize their diagnostic value, thus expanding screening to minorities that lack access to community-based ambulatory care.

[0127] Despite significant progress in Al in cardiovascular imaging, most currently available Al solutions for echocardiography have been developed for use with standard transthoracic echocardiography. Such models benefit, among others, from videos of high quality and anatomical fidelity that are obtained by certified echocardiography technicians, as well as standardized protocols with almost complete data capture that often include multiparametric phenotyping by Doppler and / or strain imaging. Unfortunately, disparities in accessing ambulatory outpatient care do exist and have repeatedly been shown to disproportionately affect marginalized communities. On this note, equitable deployment of novel Al technologies may require that these are available across all points-of-care. Indeed, as shown in our analysis, the rates of POCUS imaging among Black patients were ~3 times higher than the rate of routine transthoracic echocardiography. This further highlights the missed potential that lies in maximizing diagnostic inference from images that are already acquired as part of the current clinical workflows, particularly for conditions such as HCM, ATTR-CM and AS that benefit from early detection and risk stratification, yet remain under-diagnosed, particularly among under-represented communities.

[0128] With this unmet need in mind, the present Example describes innovative findings on both the methodological and clinical front. On the one hand, our work describes a scheme that can maximize the performance of echocardiography-based Al tools when deployed to real-world POCUS studies by addressing POCUS-specific challenges as part of the training process. We illustrate that natural augmentation of the training set through the inclusion of multiple echocardiographic views, even ones that are not included in the POCUS testing set can substantially boost model performance. By introducing a quantitative scheme to assess view quality (or anatomical correctness), we confirm that POCUS videos have a significantly lower degree of anatomical correctness compared to standard TTE studies, and conclude that naive multi-view training forces models can help adjust to different angles, depths and orientations thus improving the generalizability of Al algorithms.

[0129] On the other hand, our work has direct implications for the scalability of AI- echocardiography tools, ensuring equitable access at the first point-of-care and eliminating disparities arising from differential access and referral to outpatient testing. We demonstrate that retrospective screening of large repositories could help detect rare cardiomyopathies even though the respective images and videos were not protocolled for this purpose. This highlights the opportunity to expand opportunistic screening and phenotyping beyond dedicated echocardiographic studies, and paves the path for prospective studies testing screening of high- risk individuals in the community.

[0130] Conclusions

[0131] We have proposed and implemented a framework that enables Al-based echocardiographic algorithms to generalize to point-of-care ultrasonography, thus enabling scalable opportunistic screening of under-diagnosed cardiomyopathies.

[0132] EQUIVALENTS

[0133] Although preferred embodiments of the invention have been described using specific terms, such description is for illustrative purposes only, and it is to be understood that changes and variations may be made without departing from the spirit or scope of the following claims.

[0134] INCORPORATION BY REFERENCE

[0135] The entire contents of all patents, published patent applications, and other references cited herein are hereby expressly incorporated herein in their entireties by reference.

Claims

CLAIMSWhat is claimed is:

1. A method of training a model for cardiac phenotyping, the method comprising: providing an echocardiogram dataset; labeling the echocardiogram dataset with at least one condition of interest; splitting the echocardiogram dataset into a derivation dataset and a testing dataset; initializing a deep neural network (DNN) using random or pre-trained weights; and training a model on the derivation dataset.

2. The method of claim 1, further comprising pre-processing the echocardiogram dataset.

3. The method of claim 1, further comprising assigning a probability that a given video corresponds to a standard anatomical view with the DNN, enabling video-level classification.

4. The method of claim 1, wherein the training step further comprises at least one of introducing artificial augmentations, introducing natural augmentations, and customizing the training loss function.

5. The method of claim 4, further comprising defining a quality of the view relative to a standard set of views.

6. The method of claim 4, wherein the training step further comprises at least one of weighting for class imbalance, label smoothing, synthetic data augmentation, natural data augmentation, and loss function modification.

7. The method of claim 6, wherein the weighting for class imbalance comprises computing class-specific weights to account for an imbalance in dataset labels, the weighting for class imbalance including assigning higher weights to rare labels that are under-represented in the cohorts.

8. The method of claim 6, wherein the synthetic data augmentation comprises at least one ofrandom zero padding by a given number of pixels in each spatial dimension, random horizontal flipping, and a random rotation.

9. The method of claim 8, wherein the synthetic data augmentation accounts for at least one of variable orientation, off-axis, zoomed out, and zoomed in views.

10. The method of claim 6, wherein natural data augmentation comprises: training separate models for each view-of-interest; and training all-inclusive, view-naive models in pooled datasets; wherein the pooled datasets include all standard echocardiographic views with the classifier blinded to the input view (i.e., apical, parasternal, subcostal).

11. The method of claim 10, wherein the views-of-interest include key standard cardiac ultrasound views, routinely acquired with portable cardiac ultrasound (parasternal long axis, parasternal short axis, apical views).

12. The method of claim 6, wherein the loss function modification further comprises: weighing the loss function for the label-specific weights and view probability; wherein higher probabilities denote greater anatomical correctness compared to lower ones.

13. The method of claim 1, wherein the training step forces the model to learn generalizable signatures detectable across low and high-quality studies.

14. The method of claim 1, wherein the trained model is applicable to echocardiograms and sonograms.

15. The method of claim 1, wherein the at least one condition of interest comprises a cardiac disorder.

16. The method of claim 15, wherein the cardiac disorder comprises cardiomyopathies and valve diseases.

17. The method of claim 15, wherein the at least one condition of interest comprises transthyretin amyloid cardiomyopathy (ATTR-CM), aortic stenosis (AS), hypertrophic cardiomyopathy (HCM), and combinations thereof.

18. The method of claim 15, wherein the trained model predicts a risk of developing the at least one condition.

19. A method of cardiac phenotyping, the method comprising applying the model trained according to any one of the previous claims to one or more cardiac images.

20. The method of claim 19, wherein the one or more cardiac images include cardiac ultrasound images.

21. The method of claim 20, wherein the one or more cardiac images are obtained at the point-of-care.

22. An apparatus for cardiac phenotyping, the apparatus comprising: a processor; a memory unit; and a communication interface; wherein the processor is connected to the memory unit and the communication interface; and wherein the processor and memory are configured to implement the method of claim 19.

23. A computer readable storage medium storing computer-executable instructions for performing the method of claim 19.

Images