Adversarial deep learning model that predicts treatment effects for cardiology patients from structured electronic health record (EHR) data, aiming to forecast expected clinical outcomes of specific treatment choices given a patient's clinical status. Two autoencoders separately learn representations of patient characteristics and of the treatments given; an adversarial loss then encourages these representations to capture the correlational structure between a patient's status and the treatment received, improving downstream outcome prediction over non-adversarial baselines. Evaluated on two private cardiology EHR cohorts from a Chinese hospital, ADTEP modestly outperformed a non-adversarial ablation (DTEP) and classical baselines (logistic regression, SVM) at predicting major adverse cardiac events (MACE) after acute coronary syndrome (AUC 0.662 vs. 0.653/0.648/0.621) and at heart-failure outcome prediction.

Lightweight convolutional neural network for binary ECG classification that represents ECG beats as spectrograms (via short-time Fourier transform) rather than raw signals, after denoising and frequency filtration to reduce data volume while preserving diagnostically relevant information. Using the large public PTB-XL dataset, the spectrogram-based CNN reached 99.06% accuracy, outperforming an equivalent raw-signal CNN, while also reducing memory usage and computation by avoiding complex architectures; the authors additionally studied the effect of signal up/down-sampling on classification performance.

CNN-based arrhythmia classifier that fuses the magnitude (scalogram) and phase (phasogram) of the continuous wavelet transform (CWT) of ECG heartbeats, rather than relying on magnitude information alone as most prior 2D-representation approaches do. Several fusion strategies (input-level, intermediate-layer, and output-level fusion) were compared on the public PhysioNet MIT-BIH Arrhythmia database. Despite a simple CNN architecture, the best fusion strategy achieved about 98.5% overall accuracy, 98.5% sensitivity and 95.6% specificity, competitive with more complex state-of-the-art approaches.

Supervised contrastive learning framework for robust, fine-grained echocardiographic view classification across 47 clinically meaningful view types (rather than collapsing views into a few broad categories, as most prior work does). Introduces TTE47, the first publicly available benchmark with 47 fine-grained views independently annotated by three experts, enabling rigorous quantification of inter-observer agreement. A tailored contrastive loss produces a feature space that aligns more strongly with underlying anatomy than with any single annotator's labeling style, and the model outperforms cross-entropy and standard supervised-contrastive baselines on both TTE47 and the public TMED-2 benchmark (the latter without dataset-specific pretraining).

Self-supervised contrastive learning technique for classifying carotid plaques from ultrasound images under label scarcity. In a pretext task, a triplet network takes three augmented views (strong- and weak-augmentation) of each image and promotes their similarity from both feature- and instance-level perspectives to learn effective plaque representations; the resulting encoder is then fine-tuned on labeled ultrasound images for the downstream classification task. FeaCL achieved 83.4% classification accuracy using only 30% of the training data -- a 16.3% improvement over the same network trained without the self-supervised pretext task.

Interpretable coronary heart disease (CHD) risk prediction model based on the LightGBM gradient-boosting algorithm, combined with SHAP (SHapley Additive exPlanations) values to make individual risk predictions explainable to clinicians. Trained on the public BRFSS_2015 survey dataset and externally validated on the Framingham and Z-Alizadeh Sani datasets, the model reached 90.60% accuracy and 81.06% AUROC on BRFSS_2015, with SHAP analysis identifying age, smoking status, diabetes, hypertension, and high cholesterol as the most influential risk features. A companion CHD scoring system was built from the model to give clinicians a user-friendly risk-assessment tool.

Optimized Hierarchical Fused Fuzzy Deep Reinforcement Learning model for predicting non-sinus (higher-risk) cardiac rhythms from PQRST-analyzed 12-lead ECG data. The three-stage approach combines data preprocessing, reinforcement learning, and fuzzy deep learning to classify sinus vs. non-sinus rhythms. Evaluated on a 12-lead ECG dataset of 10,646 patients, OHFFDRL achieved 94% accuracy, an AUC of 0.91, and was interpreted using SHAP, LIME, calibration curves, adversarial vulnerability analysis, and integrated gradients; TAxis (ventricular repolarization movement range) was found to be the most important distinguishing feature.

Prototype-based neural network for interpretable, uncertainty-aware classification of aortic stenosis (AS) severity from B-mode echocardiography videos. Rather than a black-box prediction, ProtoASNet bases its output on similarity scores between the input video and a set of learned spatio-temporal prototypes (typically highlighting valve calcification and restricted leaflet motion), and uses an abstention loss to flag ambiguous/uncertain cases for expert review. Evaluated on a private clinical dataset and the public TMED-2 dataset, it achieved balanced accuracy of 80.0% (private) and 79.7% (TMED-2), improving to 82.4% when uncertain cases are excluded.

Self-supervised learning (SSL) model that identifies chronically stressed mother-fetus dyads from raw maternal abdominal ECG (aECG), which contains both maternal and fetal cardiac signals. Built on a self-supervised representation-learning approach originally developed for ECG-based emotion recognition, the model is pretrained on public ECG datasets and evaluated on a cohort of pregnant women with chronic stress exposure validated by psychological inventory, maternal hair cortisol, and the fetal stress index (FSI). Using maternal ECG alone with the publicly pretrained model, it detected the chronic-stress-exposure group with AUROC 0.982 and predicted psychological stress score (R2 0.943), FSI (R2 0.946), and maternal hair cortisol (R2 0.931).

Transformer-based multi-view multiple-instance learning (MIL) framework for patient-level coronary stenosis classification from multi-view invasive coronary angiography. Rather than requiring expensive view-level stenosis annotations, SegmentMIL is trained end-to-end on real-world clinical data using only patient-level labels already present in hospital systems, and jointly predicts stenosis presence while localizing the affected artery (left/right) and segment. It captures temporal dynamics and dependencies across the multiple angiographic views per patient (which prior view-level models ignore), and outperforms both single-view models and classical MIL baselines on internal and external clinical evaluations.

Real-time deep-learning model that fuses serial 12-lead ECG waveforms with sequential vital signs and routinely available clinical data to predict hospital admission early during emergency department (ED) encounters with cardiac presentations (chest pain, dyspnea, syncope, presyncope). Developed and validated on the public MIMIC-IV, MIMIC-IV-ED, and MIMIC-IV-ECG databases (n=30,421 ED stays with >=1 ECG; n=11,273 with >=2 ECGs), the model improves on baseline tabular (random forest) and ECG-only models by leveraging how a patient's risk evolves with successive ECGs during the visit, addressing a key limitation of single-time-point risk scores.

Pediatric-specific extension of EchoNet-Dynamic: a video-based deep learning model that segments the left ventricle and estimates ejection fraction (EF) from apical-4-chamber (A4C) and parasternal short-axis (PSAX) pediatric echocardiogram clips. Because adult-trained echo models generalize poorly to children (who vary widely in heart size, rate, and image quality), EchoNet-Peds was trained from scratch on a dedicated pediatric video dataset. It segments the LV with a Dice similarity coefficient of 0.89 in both views, estimates EF with a mean absolute error of 3.66%, and identifies pediatric systolic dysfunction with an AUC of 0.95, significantly outperforming an adult-trained model applied to the same pediatric data.

Automated deep learning workflow that detects and grades tricuspid regurgitation (TR) severity from full transthoracic echocardiography studies. The pipeline first identifies apical-4-chamber (A4C) video clips with color Doppler across the tricuspid valve from a full echo study, then applies a dedicated R(2+1)D video classifier to grade TR severity. Trained on over 2 million echo videos from 47,312 studies at Cedars-Sinai Medical Center and externally validated at Stanford Healthcare, the model identified color-Doppler A4C views with AUC ≥0.999 and detected clinically significant (moderate-or-severe) TR with AUC 0.951 and severe TR with AUC 0.980. Code and trained model weights are released to support prospective evaluation of AI-assisted TR screening.

Self-supervised deep learning system that detects ECG anomalies indirectly, by learning to forecast what a normal ECG signal should look like next. FADE is trained only on normal ECG segments using a novel morphology-inspired loss function; at inference time, a large mismatch between the forecast and the observed signal flags an anomaly, avoiding the need for labeled abnormal-beat datasets. Evaluated on the public MIT-BIH NSR and MIT-BIH Arrhythmia databases, FADE reached an average accuracy of 83.84% for anomaly detection and 85.46% for correctly classifying normal ECG, and the approach can be adapted to new recording contexts via domain adaptation.

Explainable machine learning model that detects and localizes left ventricular (LV) scar in hypertrophic cardiomyopathy (HCM) patients directly from 12-lead ECG, as a faster and cheaper alternative to late-gadolinium-enhancement (LGE) cardiac MRI, the clinical gold standard. XplainScar first uses an HCM-specific ECG segmentation algorithm to extract morphological features (duration, amplitude, slope, energy) from the QRS complex, ST segment and T wave of each lead, then combines unsupervised and self-supervised representation learning to predict scar presence and reveal which ECG features are associated with scar location (basal, mid, or apical LV). Trained on 500 HCM patients from the Johns Hopkins HCM Registry and validated on a held-out cohort of 248 HCM patients from UCSF, it reached 88% precision, 90% sensitivity, 78% specificity and an F1-score of 89% for scar detection on the external test set, analyzing a 10-patient batch of ECGs in under one minute.

Multi-label 12-lead ECG diagnosis model submitted to the PhysioNet/Computing in Cardiology Challenge 2020, built on the same residual 1D CNN family as the authors' earlier Nature Communications model but retrained and validated across the challenge's large, multi-institutional pooled training set (CPSC2018, China 12-Lead ECG Database, St. Petersburg INCART, PTB and PTB-XL, and the Georgia 12-Lead ECG Database). The model uses an unsupervised pretraining stage -- predicting unseen samples of a partially masked ECG signal -- before supervised fine-tuning to jointly detect nine diagnostic classes (atrial fibrillation, first-degree AV block, left and right bundle branch block, normal rhythm, premature atrial/ventricular contraction, and ST-segment depression/elevation). The 2020 Challenge was notable for requiring every team to publicly release both their trained model weights and full training code, making this one of relatively few 12-lead ECG classifiers with an end-to-end reproducible public pipeline.

nnU-Net-based pipeline for automated detection and sub-classification of acute thoracic aortic dissection (AD) on heterogeneous CT imaging, formulated as a semantic segmentation task rather than direct image classification. The model segments the false lumen (ascending and descending) and the dissection membrane -- along with optional indirect signs such as hemopericardium, aortic wall hematoma, and supra-aortic branch dissection -- and a patient is classified as AD-positive if at least two of the three primary segmented regions exceed a volume threshold tuned via Youden's index; the same pipeline additionally flags Stanford type A dissections. Trained on 157 heterogeneous internal CT studies (not restricted to a single contrast protocol) from Mannheim University Medical Centre and evaluated on an internal held-out test set as well as public external datasets (ImageTBAD and AVT), the model reached an AUROC of 98.7% internally and 97.0% externally, and correctly flagged 93.3% of dissection cases that had not been clinically suspected before imaging. The authors state the trained network will be made publicly available as a non-medical device for further scientific research.

Coronary artery calcium (CAC) scoring model that transfers a CNN trained for calcium scoring on non-contrast CT (NCCT) to coronary CT angiography (CCTA), where iodinated contrast otherwise confounds calcium detection and large annotated CCTA training sets are scarce. The CAC-scoring CNN is split into a feature generator and a classifier; the feature generator is trained on the NCCT source domain and adapted to the CCTA target domain via adversarial learning combined with a maximum-mean-discrepancy loss, while the source-domain classifier is reused unchanged for the target domain. Builds directly on the authors' earlier non-contrast CT calcium-scoring network.

End-to-end deep learning framework that predicts the procedural outcome of percutaneous coronary intervention (PCI) for chronic total occlusion (CTO) lesions directly from preprocedural coronary CT angiography, aiming to replace slower manual scoring systems (J-CTO, CT-RECTOR, KCCT). The pipeline first segments the coronary artery tree (Patch-UCTNet), detects candidate CTO lesions along the delineated vessel, extracts pathological lesion features with a Swin Transformer, and classifies two outcomes: successful guidewire crossing within 30 minutes and overall PCI success. In the original study, the model completed reconstruction and analysis 85% faster than manual scores (73.7s vs. 418-467s) and was more accurate than the manual CT-RECTOR, KCCT, and J-CTO_CCTA_ scores, reaching an AUROC of 0.97 on the internal test set and 0.96 on an independent external validation cohort (186 patients, 189 CTO lesions).

Two-stage convolutional neural network that automatically detects and anatomically labels coronary artery, thoracic aorta, and cardiac-valve calcifications in low-dose chest CT acquired for lung-cancer screening. A first CNN with a large receptive field (via dilated convolutions) identifies and anatomically labels candidate calcifications; a second CNN filters true positives from the candidates. Trained and evaluated on 1,744 CT scans from the National Lung Screening Trial (NLST), reaching an F1 of 0.89 (soft-filter reconstructions) / 0.84 (sharp-filter reconstructions) for coronary artery calcifications and a linearly-weighted kappa of 0.90-0.91 for per-subject cardiovascular risk categorization versus the manual reference standard.

Dual Bayesian ResNet model for heart murmur detection from multi-location phonocardiogram (PCG) recordings, developed for the George B. Moody PhysioNet Challenge 2022. Each patient's PCG recordings are segmented into overlapping log-mel spectrograms, which are passed through two Bayesian ResNet binary classifiers running simultaneously (present vs. unknown-or-absent, and unknown vs. present-or-absent); the two outputs are aggregated into a patient-level present/unknown/absent murmur classification. An optional second-stage XGBoost model integrates the DBRes output with demographic data and hand-crafted signal features. On the Challenge's official hidden test set, DBRes achieved a weighted accuracy of 0.771 for the murmur-detection task, placing 4th among all teams.

Deep learning pipeline for detection and quantification of abdominal aortic aneurysm (AAA) -- a typically asymptomatic condition often missed incidentally by radiologists -- from abdominal-pelvic CT. A modified 3D U-Net segments the aorta on both contrast and non-contrast CT volumes with a variable number of images, after which an ellipse-fitting post-processing step measures the aortic cross-sectional diameter along the vessel to detect aneurysmal dilation. Trained and validated on 321 abdominal-pelvic CT examinations from Massachusetts General Hospital, the model reached a sensitivity/specificity of 0.91/0.95 on the primary validation set, and 0.85/1.0 on a separate 57-exam generalization test set with different patient demographics and acquisition characteristics; the authors report that DeepAAA exceeded literature-reported radiologist performance for incidental AAA detection.

Early semi-supervised sequence model for estimating cardiovascular and metabolic risk directly from consumer wearable sensor data (heart rate, step count, and activity level), rather than from clinical-grade ECG or imaging. A multi-task long short-term memory (LSTM) network is first pretrained using semi-supervised sequence learning or heuristic pretraining on unlabeled wearable time series, then fine-tuned to jointly predict four self-reported conditions: diabetes, high cholesterol, high blood pressure, and sleep apnea. Trained and validated on 57,675 person-weeks of data from participants in UCSF's Health eHeart study using the Cardiogram app on Fitbit, Apple Watch, or Android Wear devices, DeepHeart outperformed hand-engineered heart-rate-variability biomarkers from the medical literature, reaching AUROCs of 0.845 (diabetes), 0.744 (high cholesterol), 0.809 (high blood pressure), and 0.830 (sleep apnea). The paper was an early demonstration that population-scale, passively-collected wearable heart-rate data could support cardiometabolic risk screening without any dedicated clinical measurement.

Clinically-informed modification of the ResNet-18 architecture for identifying occlusion myocardial infarction (OMI) -- a severe, often ST-elevation-negative heart attack caused by complete blockage of a coronary artery -- from a single 12-lead ECG. The network first learns lead-specific temporal features via 1xk temporal convolutions, then learns cross-lead spatial concordance/discordance (e.g. reciprocal ST changes) via a 12x1 spatial convolution placed after the residual blocks, with saliency maps highlighting the most relevant leads and waveform regions for explainability. Benchmarked against ResNet-18 and other CNN/random-forest baselines on a multisite real-world clinical dataset of 10,893 ECGs (OMI rate 6.5%), reaching a test AUROC of 0.889 and an average precision of 0.587, outperforming the compared models.

Ensemble classifier combining hand-engineered expert features with a deep convolutional neural network for classifying cardiac rhythm from a single-lead ECG recording into normal sinus rhythm, atrial fibrillation, another rhythm, or too noisy to classify. A large set of expert features (from time-, frequency-, and template-based analysis) is fed into a gradient-boosted tree classifier (AdaBoost), and its output is combined with a separate deep CNN operating directly on the raw waveform; combining both feature families measurably outperformed either alone. ENCASE won 1st place in the PhysioNet/Computing in Cardiology Challenge 2017 (single-lead AF classification) with an overall F1 score of 0.83 on the official hidden test set, and remains a widely cited example of combining classical signal-processing features with deep representations for ECG classification.

Video-based deep learning system for automated screening of rheumatic heart disease (RHD) from echocardiography, aimed at low-resource settings where RHD -- the most common acquired heart disease in children and young adults worldwide -- is endemic but echocardiography expertise is scarce. A 3D convolutional neural network (C3D) classifies each echo video clip, explicitly modeling temporal information across frames; a supervised meta-classifier then aggregates the per-video predictions from an exam (which may contain many video clips from different views) into a single exam-level RHD diagnosis. Evaluated on 11,646 echocardiography videos from 912 screening exams collected in underserved areas of Brazil and Uganda, the 3D C3D network significantly outperformed a comparison 2D CNN (VGG16) that ignores temporal information, and the learned aggregation model reached 72.77% exam-level diagnostic accuracy, exceeding simple majority voting across a patient's videos.

Heart murmur detection model combining spectrogram-derived deep features with hand-crafted peak-interval (PI) features extracted from phonocardiogram recordings, submitted to the George B. Moody PhysioNet Challenge 2022 (team CAU_UMN) and later extended into the 'SpectroHeart' method. Peak-to-peak interval sequences and their summary statistics are combined with spectrogram representations of the PCG signal, optionally alongside patient demographic data, to classify murmur presence across multiple auscultation locations. The team's Challenge submission placed 5th of all teams on the murmur-detection task.

U-Net-variant segmentation model that identifies and quantifies coronary artery calcium (CAC) directly from routine non-gated, non-contrast chest CT scans -- the kind ordered for lung-cancer screening or unrelated indications rather than a dedicated cardiac scan -- so that the tens of millions of such scans performed annually can be opportunistically screened for cardiovascular risk without any extra imaging. Predicted calcium masks are combined with the CT's Hounsfield units to compute an Agatston-equivalent score. Trained on 446 expert-segmented scans from 98 medical centers across the U.S. Department of Veterans Affairs national health system (capturing substantial heterogeneity in scanners and protocols) and benchmarked against 795 patients with a paired same-year gated CAC study: nongated AI-CAC differentiates zero-vs-nonzero and <100-vs->=100 Agatston categories with 89.4% (F1 0.93) and 87.3% (F1 0.89) accuracy respectively, and its score stratifies 10-year all-cause mortality (CAC 0 vs. >400: 25.4% vs. 60.2%, hazard ratio 3.49) and composite stroke/MI/death risk (33.5% vs. 63.8%, hazard ratio 3.00). In a simulated opportunistic-screening run across 8,052 low-dose CT scans, cardiologists confirmed 99.2% of patients flagged with AI-CAC >400 would benefit from lipid-lowering therapy. Code and trained model weights are both public under an MIT license.

Unified deep-learning model for 12-lead ECG analysis that predicts a broad range of cardiac and non-cardiac discharge diagnoses coded under the ICD-10 classification system, evaluated as a unified screening tool for emergency departments where a single ECG could flag many potential conditions at once rather than one disease at a time. Introduces the MIMIC-IV-ECG-ICD-ED benchmark dataset (derived from MIMIC-IV and MIMIC-IV-ECG) and reports AUROC scores across diverse diagnostic scenarios (all discharge diagnoses vs. emergency-department-only diagnoses, cardiac vs. non-cardiac ICD-10 chapters), suggesting integration into emergency-department clinical decision-support systems.

Multi-outcome AI-ECG risk-estimation platform that turns a single 12-lead ECG into a patient-specific survival curve, using a residual convolutional neural network trained with a discrete-time survival loss to predict not just risk but time-to-event. Beyond all-cause and cardiovascular mortality, separately fine-tuned heads predict future ventricular arrhythmia, complete heart block, atrial fibrillation, atherosclerotic cardiovascular disease, heart failure, and (in follow-up work) hypertension -- all from a single resting ECG, including in ECGs a cardiologist would read as normal. Derived on 1.16 million ECGs from 189,539 Beth Israel Deaconess Medical Center patients and externally validated across transnational cohorts in the USA, Brazil (CODE), and the UK (UK Biobank). A variational autoencoder and genome/phenome-wide association analyses were used to show the model's predictions track biologically plausible features (QRS morphology, LV structure/function, and loci linked to cardiac structure, QT interval, and biological aging). NHS trials of AIRE were planned for late 2025. No public code or model weights have been released.

Multimodal machine learning model that flags patients at risk of transthyretin amyloid cardiomyopathy (ATTR-CM) -- a progressive, underdiagnosed disease with expanding disease-modifying treatment options -- from routinely available ECG waveforms, echocardiographic measurements, demographics, and diagnosis codes for orthopedic manifestations of amyloidosis (e.g. carpal tunnel syndrome, spinal stenosis). Developed on 799 patients with 5-fold cross-validation (AUROC 0.85) and externally validated on 422 patients at a separate site (AUROC 0.82), with consistent accuracy across Hispanic, non-Hispanic Black, and non-Hispanic White patients. In a subsequent nonrandomized, single-system, multisite clinical trial (the Cardiac Amyloidosis Discovery Trial), patients flagged by ATTRACTnet and referred for confirmatory amyloid scintigraphy were positive for ATTR-CM 48% of the time, more than 2.8x the positivity rate of historical (15.3%) and contemporary (17.0%) controls referred by usual clinical judgment (P < .001 for both). This is a proprietary clinical AI program; no public code or model weights have been released.

ECG-guided photoplethysmography (PPG) foundation model pretrained on over 100,000 hours of synchronized PPG-ECG recordings from 58,796 subjects across five clinical and wearable sources, using a CLIP-style contrastive alignment framework so the PPG encoder inherits physiologically grounded structure from paired ECG. Achieves state-of-the-art performance on 13 of 15 conventional physiological-analysis tasks across eight datasets, and shows meaningful discriminative capability (AUC >= 0.70) for 307 ICD-10-coded phenotypes across 16 phecode chapters, including many non-cardiovascular conditions.

Foundation vision-language model for cardiac MRI that learns pathophysiological visual representations directly from the natural-language radiology reports accompanying each scan, rather than from hand-labeled targets. A Multi-scale Vision Transformer (MViT, Kinetics-400-initialized) video encoder for cine CMR sequences is contrastively pretrained (InfoNCE) against a PubMed-pretrained BERT text encoder over 19,041 multi-institutional CMR studies. The frozen vision encoder transfers with strong performance to left-ventricular ejection-fraction regression (MAE 3.34% on a UK Biobank hold-out of ~4,259-45,623 participants) and detecting HFrEF (LVEF<40%, AUC 0.880), and the paper reports emergent zero-/few-shot performance across 39 cardiac and non-cardiac conditions including cardiac amyloidosis and hypertrophic cardiomyopathy. Code and pretrained MViT encoder weights are both released (Hugging Face, CC BY-NC 4.0).

Vision Transformer-based classifier that detects abnormal left-ventricular ejection fraction (LVEF < 50%) directly from static, non-gated, non-contrast chest CT scans -- an imaging modality ordered for unrelated indications (lung cancer screening, pulmonary embolism, trauma) in over 80 million US exams a year -- as a form of opportunistic heart-failure screening. Fine-tunes the encoder of the CT-ViT (GenerateCT) framework, with separate spatial (axial-plane) and z-axis (slice-wise) self-attention blocks, on 3D CT volumes paired with echocardiogram-derived LVEF labels from 25,948 Columbia University studies; reaches an AUROC of 0.786 on a held-out test set and 0.762 on external validation at Weill Cornell Medicine, clearly outperforming demographic/diagnosis-code-only baselines (Random Forest, XGBoost, AUROC 0.54-0.61). On a radiologist-comparison subset, the model's weighted F1 (0.80-0.81) exceeded two board-certified thoracic radiologists (0.62-0.80) at a small fraction of the interpretation time. Grad-CAM saliency maps highlighted clinically sensible correlates of reduced LVEF (cardiomegaly, dilated superior vena cava, calcified ascending aorta, pacemaker hardware, pulmonary edema). No public code or model weights have been released.

Multimodal deep-learning framework that estimates and forecasts abnormal laboratory values directly from a 12-lead ECG plus routinely available demographics, biometrics, and vital signs -- reframing dozens of blood tests as binary classification targets predictable from a test that is already fast, non-invasive, and nearly universal in acute care. A structured state-space (S4) encoder processes the raw ECG waveform and is late-fused with an MLP encoder over the tabular metadata; the same architecture is trained both to estimate the closest lab value within 60 minutes of the ECG ('abnormality prediction') and to forecast whether a value will become abnormal 30/60/120 minutes into the future ('abnormality forecasting'). Trained and evaluated on 385,480 linked ECG-lab-value samples from 127,994 MIMIC-IV patients, the model reaches AUROC > 0.7 for 24 distinct lab abnormalities in the prediction setting and 24 in the forecasting setting, spanning cardiac, renal, hematological, metabolic, immunological, and coagulation categories -- with NT-proBNP elevation the best-predicted marker (AUROC 0.90), followed by hemoglobin, albumin, and hematocrit derangements (AUROC > 0.82). Code for dataset construction, training, and evaluation is public under an MIT license; no pretrained model weights are released.

Fully automated pipeline for interpreting coronary angiograms that chains four purpose-built neural networks: (1) angiographic projection-angle identification, (2) left/right coronary artery detection, (3) arterial segment localization, and (4) stenosis-severity estimation. Trained on 13,843 angiographic studies (195,195 videos) from 11,972 adult patients at UCSF (2008-2019), with projection-angle and LCA/RCA-detection tasks each reaching precision/sensitivity/F1 at or above 90%. For predicting obstructive coronary artery disease (>=70% stenosis), CathAI reaches an AUC of 0.862 internally, 0.869 on external angiograms from the University of Ottawa Heart Institute, and 0.775 after retraining on quantitative-coronary-angiography labels from the Montreal Heart Institute core lab. No public code or model weights have been released.

Multi-view foundation model for coronary angiography trained with video-text contrastive learning on 203,808 angiography videos from 28,117 patients across 32,473 studies at the Montreal Heart Institute, externally validated on 4,249 studies from UCSF. Integrates multiple angiographic projections with attention-based pooling for study-level assessment spanning diagnostic, prognostic, and disease-progression tasks: significant-stenosis detection (AUROC 0.888 internal / 0.89 external), stenosis-percentage estimation (MAE 13.6% vs. 19.0% for clinical reports), chronic total occlusion, intracoronary thrombus, and coronary calcification detection. Transfer learning further enables one-year MACE prediction (AUROC 0.79) and LVEF estimation (MAE 7.3%) from the same angiography embeddings, with a mean in-hospital inference time of 4.2 seconds.

AI-driven pipeline for quantitative coronary-stenosis assessment from routine DICOM coronary angiography videos, combining vessel tracking with a video Swin3D transformer trained and validated on 182,418 angiography videos spanning 5 years at the Montreal Heart Institute. Achieves a mean absolute error of 20.15% and a classification AUROC of 0.8294 for stenosis-percentage prediction against cardiologist assessment, with lower inter-rater variability than two expert interventional cardiologists, and can be fine-tuned to quantitative coronary angiography (QCA) data for even lower error (MAE 7.75%).

Supervised EfficientNetV2-based 12-lead ECG model trained on over 1 million ECGs from the Montreal Heart Institute to predict 77 cardiac conditions derived from American Heart Association recommendations, plus fine-tuned digital-biomarker heads for reduced LVEF, 5-year atrial-fibrillation risk, and long-QT-syndrome (LQTS) detection/genotyping. Validated on 881,403 ECGs across 11 geographically diverse cohorts (4 public, 7 private health systems), achieving AUROCs above 0.98 for the 77-condition interpretation task while being 60x smaller and 29x faster at inference than its self-supervised DeepECG-SSL counterpart, with up to 9.7x lower CO2 emissions on equivalent tasks.

Self-supervised EfficientNetV2-based 12-lead ECG foundation model pretrained via contrastive learning and masked-lead modeling on 1.9 million ECGs (Montreal Heart Institute plus CODE-15% and MIMIC-IV), then fine-tuned for the same 77-condition ECG interpretation task and digital-biomarker extraction as DeepECG-SL. Outperforms the supervised counterpart on label-scarce digital-biomarker tasks, with the largest gains on LQTS genotype classification (AUROC 0.931 vs. 0.850, n=127 ECGs) and 5-year atrial-fibrillation risk (AUROC 0.742 vs. 0.734, n=132,050 ECGs), and outperforms ECG-FM and ECGFounder on shared external diagnostic classes.

Open-weight video-based deep neural network that predicts reduced right-ventricular systolic function (RVSF) directly from routine left and right coronary angiogram videos, enabling real-time RV-dysfunction screening in the catheterization lab when echocardiography is unavailable. Built on an X3D-M spatiotemporal video architecture (Kinetics-400 pretrained) that aggregates per-video probabilities into a study-level normal-vs-reduced RVSF classification, with Grad-CAM/Guided-Backpropagation explainability confirming attention to RV-specific coronary motion rather than left-ventricular signal. Trained on 8,053 angiographic studies from 6,923 Montreal Heart Institute patients (2017-2023), externally validated at UCSF, and prospectively deployed at MHI via the PACS-AI platform, where AI assistance improved reader accuracy from 72.1% to 77.6% for cardiologists and 43.5% to 64.0% for medical students.

Multimodal large language model for ECG medical-report generation and cardiology conversational question-answering. An ECG-CoCa encoder (contrastive ECG-report pretraining in the style of OpenCLIP) is paired with a LLaVA-style vision-language architecture and an LLM backbone, fine-tuned on a purpose-built 45k-example ECG-instruction dataset (19k diagnosis examples + 25k multi-turn dialogue examples) built from five public 12-lead ECG datasets. Produces free-text diagnostic reports and supports zero-shot ECG-report retrieval classification.

Self-supervised ECG representation-learning method that adapts Joint-Embedding Predictive Architecture (JEPA) -- originally developed for images -- to 1D electrocardiogram signals. A Vision Transformer encoder (ViT-XS/S/B) is pretrained to predict masked temporal segments of the ECG directly in latent feature space, using a masking strategy tailored to time-series, on more than 1 million ECGs pooled from MIMIC-IV-ECG, CODE-15%, PTB-XL, Chapman-Shaoxing, CPSC2018/Extra, Georgia, PTB, and St-Petersburg-INCART. After fine-tuning on PTB-XL, the ViT-S/JEPA model reaches 0.945 AUC on the all-statements diagnostic task, exceeding prior self-supervised ECG baselines including CPC and ST-MEM. Developed at the Zuse Institute Berlin.

Explainable Inception-style 1D CNN for multi-label arrhythmia detection from 12-lead ECGs, integrating Grad-CAM visualization to highlight the waveform segments driving each prediction. Trained on MIMIC-IV-ECG and externally validated on PTB-XL across atrial fibrillation, sinus tachycardia, conduction disturbances (RBBB/LBBB/LAFB), long QT, Wolff-Parkinson-White pattern, and paced-rhythm detection, with all metrics exceeding 90% internally and strong generalization on external validation.

Multi-instance contrastive self-supervised learning framework for echocardiography video representation learning: two distinct videos from the same patient exam are treated as positive pairs (rather than augmentations of a single clip), and a frame-reordering pretext task on temporally shuffled frames encourages the 3D-CNN backbone to learn temporally coherent representations. After self-supervised pretraining on unlabeled echocardiograms, the backbone is efficiently fine-tuned for cardiac disease classification (severe aortic stenosis, left ventricular hypertrophy) using very few labeled examples, outperforming SimCLR, standard multi-instance SimCLR, and Kinetics-400-initialized baselines across a range of training-data ratios.

Predicts cardiac MRI (CMR) tissue-characterization findings -- wall-motion abnormalities and myocardial scar -- directly from standard transthoracic echocardiography videos (A4C/A2C/PLAX views), using a factorized 3D R2+1D convolutional network. Trained and validated on a single-institution cohort of more than 1,400 patients with paired echo and CMR studies within 30 days of each other. Released weights and inference code cover the two binarized outcomes (wall motion, scar); the continuous CMR tissue markers (native T1, T2, ECV) evaluated in the paper are not part of the public release.

First multimodal LLM to unify ECG time series, 12-lead ECG images, and text for grounded, clinician-aligned ECG interpretation. A dual-encoder framework (ECG-CoCa time-series encoder plus a LLaVA-style vision-language backbone) extracts complementary time-series and image features with cross-modal alignment, trained on knowledge-guided instruction data (ECG-Grounding, linking diagnoses to measurable waveform parameters such as QRS/PR intervals) plus the 1.15-million-conversation ECG-Instruct corpus. Introduces the "Grounded ECG Understanding" benchmark and improves predictive performance, explainability, and grounding over prior ECG-language models such as ECG-Chat and PULSE.

Multi-scale ECG-language pretraining model that aligns 12-lead ECG signals with clinical text reports at three granularities -- token, beat, and rhythm level -- rather than a single global embedding. First fine-tunes a cardiology-specialized text encoder to improve understanding of ECG report language, then trains an ECG-FM-initialized ECG encoder against it with hierarchical contrastive supervision. Outperforms prior ECG-language and self-supervised baselines including MERL, ST-MEM, and HeartLang on zero-shot classification, linear probing, and ECG report generation, with especially large gains at low label fractions. Developed at the University of Hong Kong (HKU-MedAI).

Multi-instance-learning (MIL) model for detecting >=50% coronary stenosis directly from curved multiplanar reformation (CMR) images generated during routine coronary CT angiography (CCTA) reads, without requiring slice-level annotations. A multi-range Hounsfield-unit preprocessing pipeline (Sobel edge detection across five attenuation windows) highlights plaque and vessel-wall structures, which a VGG16-based encoder with positional encoding and multi-head attention aggregates across each patient's 'bag' of up to 36 CMR slices per artery to give an interpretable, attention-weighted patient-level prediction. Trained and five-fold cross-validated on 900 real-world CCTA cases (776 LAD / 694 RCA / 600 LCX) from Sahlgrenska University Hospital, reaching AUCs of 0.91-0.92 across the three major coronary arteries. Code (preprocessing + MIL training pipeline) is public; the clinical CMR dataset and trained weights are not released.

Ensemble of 3D convolutional neural networks that detects severe aortic stenosis directly from single-view 2D parasternal long-axis (PLAX) transthoracic echocardiogram videos, without requiring Doppler imaging. Representations are first pretrained with patient-level contrastive self-supervised learning on PLAX clips, then fine-tuned for binary AS classification; the ensemble is externally validated across a temporally-distinct cohort and geographically-distinct cohorts in California and New England. No pretrained weights are released; only the training/evaluation pipeline is public.

Patch-based masked-training framework for robust arrhythmia detection from digitized, multi-layout ECG images (e.g. 3x4, 2x6, 12x1 printed/scanned layouts), designed to handle the asynchronous lead timing and partial signal blackout that digitization introduces. An adaptive variable block-count masking strategy focuses model attention on key patches with cross-lead dependencies. Evaluated on PTB-XL digitized into multiple synthetic layouts and externally validated on 400 real digitized ECG images from Chaoyang Hospital, outperforming classical imputation baselines and the CNN foundation model ECGFounder.

Self-supervised foundation model for photoplethysmography (PPG) that explicitly targets the low-signal-quality problem endemic to real-world wearable and bedside PPG, rather than assuming clean input. A Siamese ResNet (SimSiam-style) encoder is pretrained to produce similar representations for temporally-adjacent high- and low-quality PPG segments presumed to share the same underlying physiological state, with curriculum learning that gradually increases the artifact-level gap between paired segments. Pretrained on over 36 million 30-second PPG pairs drawn from more than 24,100 UCSF ICU patients, then fine-tuned and evaluated on six downstream cardiovascular-monitoring datasets/tasks (heart rate, respiratory rate, blood pressure, and atrial-fibrillation detection), where it exceeds prior state-of-the-art by roughly 75% on respiratory-rate estimation and 5% on AF detection, with performance scaling with backbone depth (ResNet152 best). Code and the pretraining dataset are not public; the pretraining data is available upon request under a UCSF-Emory data use agreement.

ECG foundation model designed to remain accurate when leads are missing or signals are noisy. A 1D ConvNeXt V2 encoder is trained with a dual-mode self-distillation objective (separate lead-missing and noise "teachers") alongside contrastive alignment to detailed diagnostic-criteria text reports retrieved via a lightweight, LLM-free "Cardiac Feature Retrieval" module. Consistently ranks best or second-best across PTB-XL diagnostic tasks and MIT-BIH arrhythmia classification under original, noisy, lead-missing, and combined-corruption conditions. Developed by FPT Software AI Center and the University of Arkansas.

Multimodal cardiac MRI foundation model that fuses 3D+T cine CMR (short-axis and long-axis views) with tabular patient health records (demographics, metabolic, and lifestyle factors) from 42,000 UK Biobank participants. Two-stage self-supervised pretraining -- masked-image reconstruction, then imaging-tabular contrastive alignment -- produces representations that transfer to whole-heart segmentation, cardiac phenotype/physiological-feature regression, and cardiac/metabolic disease classification within one unified framework.

Foundation model that encodes single-lead (lead I) ECGs with information from paired transthoracic echocardiography reports, aimed at label-efficient screening for structural heart disease (SHD) on wearable and portable single-lead ECG devices. A 7-layer 1D-CNN ECG encoder and a RoBERTa-based text encoder are contrastively pretrained (CLIP-style) on 194,551 ECG-echo report pairs from 77,378 adults in the Yale New Haven Health System, then the ECG encoder is fine-tuned on a temporally-distinct cohort to detect reduced LVEF, diastolic dysfunction, and a composite SHD label. Matches a randomly-initialized CNN at full training-data volume but substantially outperforms it in label-scarce regimes (e.g. with only 0.5% of labeled data).

ECG foundation model built on the xLSTM (extended LSTM) architecture: a bidirectional stack of nine alternating scalar- and matrix-memory LSTM blocks that scales linearly with sequence length, unlike the quadratic cost of transformer-based ECG models. Pretrained with SimDINOv2, a coding-rate-regularized self-distillation (DINO) objective adapted from computer vision to ECG time series, on roughly 8 million recordings from CODE, INCART, and Chapman-Shaoxing-Ningbo. Introduced alongside BenchECG, a standardized 8-dataset/10-task benchmark, on which xECG achieves the best average rank of any publicly available ECG foundation model, with particular strength on long-context tasks (30-minute ambulatory arrhythmia classification, multi-hour sleep-apnea segmentation) where transformer-based models are computationally limited.

Open-source, reproducible pipeline for representing 12-lead ECGs as explicit graphs -- nodes per lead-timepatch, with edges encoding established inter-lead spatial relationships (fully-connected limb- and chest-lead subgraphs bridged via leads I, aVF, V4, and V5) -- and classifying them with a Graph Convolutional Network, paired with GNNExplainer to surface which leads and lead-pairs drove each prediction. Evaluated on PTB-XL for five-class diagnostic superclass classification (AUC 0.86) and, with the same architecture, on anteroseptal-vs-inferior myocardial-infarction localization (AUC 0.92), externally validated on the population-based SHIP cohort (AUC 0.87). Explainability analysis showed the GNN's lead attention recovers standard ECG diagnostic criteria (e.g. V1-V3 for anteroseptal MI, II/III/aVF for inferior MI). Developed at University Medical Center Gottingen; code and an example trained checkpoint are released under CC BY-NC 4.0.

Video-based deep learning model that grades aortic regurgitation (AR) severity—none/trace, mild, moderate, or severe—from color Doppler echocardiography. View-specific R(2+1)D 3D-CNNs analyze five standard transthoracic views (PLAX, PLAX-AV, A3C, A3C-AV, A5C) and their outputs are combined by a maximum-severity rule at the study level. Trained on ~47,600 color Doppler videos from Cedars-Sinai and externally validated at Stanford Healthcare, reaching AUCs of 0.95 for at-least-moderate AR and 0.97 for severe AR internally. Developed by the Ouyang lab at Cedars-Sinai Medical Center.

Open-source pipeline that classifies aortic stenosis (AS) severity from transthoracic echocardiography by combining structural and functional information. Video-based R(2+1)D convolutional networks read six B-mode and color Doppler views while a segmentation model measures peak aortic-jet velocity, and an ensemble integrates these into a final severity prediction. Trained on 210,193 images from Kaiser Permanente Northern California and validated across held-out, temporally distinct, and external Stanford and Cedars-Sinai cohorts, reaching AUCs up to 0.96–0.99 for severe AS. Developed by the Ouyang lab.

Video-based deep learning model that estimates 14 common blood biomarkers and laboratory values—including hemoglobin (anemia), B-type natriuretic peptide (BNP), troponin I, and blood urea nitrogen (BUN)—directly from apical-4-chamber echocardiogram videos. Built on a spatiotemporal convolutional network (R(2+1)D-style) with residual connections that produces beat-by-beat estimates for both regression and abnormality classification. Trained on over 70,000 echocardiograms from Stanford Healthcare and externally validated at Cedars-Sinai, reaching AUCs around 0.80–0.86 for detecting anemia and elevated BNP. Developed by the Ouyang and Zou labs at Stanford University and Cedars-Sinai.

Deep learning model for automated phenotyping of mitral stenosis (MS) from echocardiography. Uses video-based R(2+1)D convolutional networks on color Doppler and B-mode views to identify and grade mitral stenosis severity, following the multi-view valvular-assessment approach of the EchoNet family. Trained and validated on large clinical echocardiography cohorts from Kaiser Permanente Northern California with external testing at Stanford Healthcare and Cedars-Sinai. Developed by the Ouyang lab.

CNN-based image classifier that screens 12-lead ECG images for a signature of transthyretin amyloid cardiomyopathy (ATTR-CM), producing a study-level probability score. Trained on a private Yale New Haven Health System cohort of nuclear-imaging-confirmed ATTR-CM cases and matched controls. Used alongside a companion echocardiography model to track pre-clinical ATTR-CM progression years before it would otherwise be confirmed by nuclear amyloid imaging. Developed by Yale's CarDS Lab and distributed as a packaged research-use executable rather than downloadable weights.

EfficientNet-B3 CNN that detects hypertrophic cardiomyopathy directly from images of printed or scanned 12-lead ECGs, rather than from raw digital waveforms, enabling screening from a photo of a paper tracing. Initialized via self-supervised contrastive pretraining on patient identity, then fine-tuned at Yale New Haven Hospital on over 124,000 ECGs from about 67,000 patients, with HCM status confirmed by cardiac MRI or echocardiography. Externally validated on ECG images from MIMIC-IV, Amsterdam UMC, and UK Biobank. Developed by Yale's CarDS Lab.

Screens 12-lead ECGs for Chagas cardiomyopathy by first pretraining a feature extractor to predict blood-biomarker levels from MIMIC-IV-ECG data, then fine-tuning on Brazilian CODE-15%, SaMi-Trop, and PTB-XL recordings; the final model is a 5-model ensemble. Submitted to the George B. Moody PhysioNet Challenge 2025 (Detection of Chagas Disease from the ECG), where it placed 5th on the official leaderboard. Developed by a team from Akershus University Hospital and the University of Oslo.

1D residual neural network that screens 10-second, 12-lead ECG tracings for six common abnormalities: first- and second-degree AV block patterns, right and left bundle branch block, sinus bradycardia, atrial fibrillation, and sinus tachycardia. Reported F1 scores above 80% and specificity over 99% when benchmarked against cardiology residents. Developed at Universidade Federal de Minas Gerais and trained on the large Brazilian CODE-15% ECG dataset.

Vision-language model that jointly embeds a cardiac MRI study, treated as video, with the impression section of its clinical report. Combines a video encoder over cine/LGE frame sequences with a Bio+ClinicalBERT text encoder using CLIP-style contrastive training. Supports zero-shot and few-shot classification of cardiomyopathies, amyloidosis, and LV dysfunction, plus image/report retrieval and structured report drafting. Trained on a private, single-institution corpus of roughly 11,000-14,000 CMR study-report pairs from Cleveland Clinic and Case Western.

Foundation model for cine cardiac MRI, self-supervised via masked autoencoding on nearly 75,000 UK Biobank scans. Uses a Vision Transformer with a convolutional stem, unified across long-axis and short-axis views. Fine-tuned checkpoints are released for ventricle and myocardium segmentation, ejection-fraction regression, cardiovascular disease classification, and landmark localization across several public benchmark datasets (ACDC, M&Ms, M&Ms2, EMIDEC, and others).

Multi-task model that jointly assesses signal quality and detects atrial fibrillation from wrist-worn wearable photoplethysmography (PPG), pretrained on roughly one million simulated unlabeled signals before fine-tuning on labeled wearable data. Uses a 1D CNN with separate output heads for signal quality and arrhythmia detection. Developed by Stanford's Ashley Lab.

Dual-encoder single-lead ECG classifier for atrial fibrillation detection that fuses a raw-signal branch with a spectrogram branch via axial attention and a Transformer. Originally developed as a graduate-course project at TU Darmstadt for the 2017 PhysioNet/CinC Challenge, and later extended in a 2023 follow-up study. Released in four sizes up to 130M parameters (S/M/L/XL), alongside a simpler CNN+LSTM variant.

Open ECG foundation model with 90.9M parameters, built on a wav2vec 2.0-style Transformer and pretrained on 1.25-1.5 million ECGs using a hybrid contrastive-and-generative self-supervised objective. Base pretrained weights and MIMIC-IV-ECG-finetuned downstream checkpoints are both released. Developed on the fairseq_signals framework by the University of Toronto / Vector Institute's Wang lab.

Multi-task 12-lead ECG model with output heads for incident atrial-fibrillation risk (as a survival curve), incident mortality risk, prevalent AF classification, sex classification, and age regression. Built on a 1D CNN over the raw waveform, and developed by the Broad Institute's ML4H group as a successor to their ECG-AI model published in Circulation. Trained on ECGs from UK Biobank and Massachusetts General Hospital, neither of which is publicly released.

Large-scale ECG foundation model pretrained on more than 10 million recordings spanning 150 label categories from the Harvard-Emory ECG Database. Built as a general-purpose feature extractor that can be fine-tuned for arrhythmia detection, demographic inference, and event prediction, and externally validated on MIMIC-IV-ECG and PTB-XL. Also used as the pretrained backbone for downstream clinical models such as Pocket-K, a hyperkalemia detector. Developed by Peking University and Harvard-Emory researchers.

Multimodal ECG model that pairs a 1D ConvNeXtV2 signal encoder with a BioLinkBERT text encoder, trained with a joint contrastive-and-captioning objective using LLM-generated descriptions of ECG demographics and waveform patterns in place of raw clinical reports. Validated on arrhythmia diagnosis and ECG-based subject identification, reaching an AUROC of 0.938 fine-tuned and 0.812 zero-shot on PTB-XL diagnostic classification. Developed at Rice University.

Echocardiography foundation model trained with a latent-predictive (V-JEPA2-style) self-supervised objective rather than pixel reconstruction, pretrained on 18 million echocardiograms from 300,000 patients drawn from the public MIMIC-IV-ECHO dataset plus a private multi-site archive - reportedly the largest echo pretraining corpus assembled to date. With a frozen backbone and only lightweight added layers, it outperforms prior echo foundation models by roughly 20% on ejection-fraction estimation and 17% on right-ventricular pressure estimation, reaches strong view-classification accuracy using just 1% of labels, and transfers zero-shot to pediatric echo better than fully fine-tuned baselines. Developed by the University of Toronto's Bo Wang Lab.

End-to-end pipeline for apical-4-chamber echocardiogram videos that segments the left ventricle, estimates ejection fraction on a beat-to-beat basis, and classifies cardiomyopathy with reduced ejection fraction. Combines a DeepLabV3-ResNet50 segmentation model with a 3D CNN (R2+1D/R3D/MC3) initialized on the Kinetics-400 video dataset. Trained on the public EchoNet-Dynamic dataset released alongside it, and one of the most widely reused open echocardiography models since its 2020 Nature publication. Developed by Stanford University.

Measures interventricular septum thickness, LV internal diameter, and posterior wall thickness from PLAX echocardiogram videos, then classifies the underlying cause of left ventricular hypertrophy as either cardiac amyloidosis or hypertrophic cardiomyopathy. Combines an atrous-convolution 2D CNN for wall-thickness segmentation with a 3D residual CNN for etiology classification. Trained on 28,201 videos across Stanford, Cedars-Sinai, and the Unity Imaging Collaborative. Developed by Stanford University.

Fully automated pipeline that scans a complete transthoracic echocardiogram study, identifies the apical-4-chamber color-Doppler clips showing the mitral valve, and grades mitral regurgitation severity at the study level. Combines a view/valve-presence classifier with a spatiotemporal CNN for severity classification. Trained on a private Cedars-Sinai cohort of 58,614 studies and externally validated on 915 studies from Stanford Healthcare.

Detects 12 categories of echocardiogram-confirmed structural heart disease from 12-lead ECG waveforms combined with demographic and clinical covariates. Uses the same architecture as the original, larger EchoNext model but is trained entirely on the public EchoNext-Mini dataset - 100,000 de-identified ECGs from Columbia University Irving Medical Center released on PhysioNet - making it one of the more fully reproducible models of its kind, with public weights, a Docker image, and inference code.

Treats ECGs as a language: a QRS-Tokenizer converts raw waveforms into discrete heartbeat 'words' from a learned 8,192-entry vocabulary, and a spatio-temporal transformer (ST-ECGFormer) is pretrained via masked-sentence modeling over these tokens. Evaluated for robust, competitive performance across six public ECG datasets and published at ICLR 2025. Developed by Peking University's digital health group, pretrained on MIMIC-IV-ECG.

Self-supervised foundation model for 12-lead ECGs, pretrained on 9.1 million recordings covering 164 cardiovascular conditions across adult and pediatric cohorts, including single-lead settings. Uses a HuBERT-style Transformer encoder and can be fine-tuned with a simple output layer for diagnosis and event-prediction tasks. Released in small, base, and large (~183M parameter) configurations by researchers at the University of Brescia.

Multimodal model that learns a shared representation space for ECG signals and their clinical text reports, pretrained on paired MIMIC-IV-ECG recordings and reports. Supports zero-shot ECG classification via text prompts, tested across six public benchmark datasets including PTB-XL and CPSC2018 without any downstream training data. Developed at Imperial College London and published at ICML 2024.

CLIP-style model that aligns 12-lead ECG signals with free-text echocardiography reports for zero-shot detection of structural heart disease directly from an ECG. Extends the MERL framework, and was trained on 45,016 paired ECG-echo reports from two Hong Kong hospitals, with external validation on the public EchoNext dataset from Columbia University. Developed by researchers at the University of Hong Kong and Imperial College London; described in a 2025 medRxiv preprint.

Reference benchmark suite for the PTB-XL ECG dataset, providing pretrained xresnet1d, InceptionTime, LSTM, and CPC-pretrained models for 71-label, diagnostic, sub-diagnostic, and super-diagnostic classification of ECG findings. Widely used as a standardized baseline for comparing new ECG classification methods. Developed by Strodthoff et al. at Fraunhofer HHI.

View-agnostic, multi-task model that performs 39 different echocardiographic reporting tasks - covering chamber size and function, valve disease, and more - from any combination of views, aggregating clip-level predictions up to the study level. Combines a ConvNeXt-Tiny frame encoder with a temporal Transformer and separate output heads per task. Trained on private Yale-New Haven Health System echo videos and published in JAMA in 2025 by Yale's CarDS Lab.