3D convolutional autoencoder that filters reverberation clutter artifacts from transthoracic echocardiography (TTE) video sequences, improving downstream measurements such as speckle-tracking strain. Built on a 3D U-Net-style encoder-decoder with an input-output skip connection to preserve fine structures and attention-gate modules to focus on cluttered regions, the network was trained on synthetic clutter simulated across six ultrasound vendors and generalized well to real in vivo artifactual sequences, substantially reducing the discrepancy between cluttered and clutter-free strain profiles while running in a fraction of a second per sequence.

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.

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 ECG embedding model inspired by BERT/RoBERTa from natural language processing, designed for efficient medical signal analysis. HeartBERT translates ECG signals into an intermediate synthetic 'language' via signal quantization and discretization (Lloyd-Max quantization), then trains a RoBERTa-style encoder from scratch on this text-like representation using the MIT-BIH Arrhythmia Database, PTB-XL, and European ST-T Database. The resulting embeddings are evaluated on two downstream tasks -- sleep-stage classification and heartbeat classification -- using bidirectional LSTM heads, showing particular strength when only small labeled training datasets are available.

Deep learning strategy for cost-effective, comprehensive cardiac screening from ECG alone, by transferring domain-specific structural information from cardiac magnetic resonance (CMR) imaging into ECG representations. Combines multimodal contrastive learning with masked data modelling during pretraining on paired ECG-CMR data, then uses only ECG at inference. On 40,044 UK Biobank subjects, the multimodal pretraining improved subject-specific CVD risk prediction by up to 12.19% and cardiac phenotype prediction by up to 27.59% versus ECG-only baselines, with learned ECG representations shown to incorporate information from CMR regions of interest.

Self-supervised deep learning model that extracts cardiovascular-risk-relevant patterns from multimodal polysomnography (PSG) signals -- EEG, ECG, and respiratory signals -- without relying on manual sleep-stage annotations. Trained on 4,398 participants, the model derives 'projection scores' by contrasting embeddings from individuals with and without cardiovascular disease (CVD) outcomes. Externally validated in an independent cohort of 1,093 participants, ECG-derived projection scores were predictive of prevalent and incident cardiac conditions (particularly CVD mortality), and combining projection scores with the Framingham Risk Score consistently improved prediction (AUC 0.607-0.965 internally, 0.710-0.807 externally across most outcomes).

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.

Self-supervised representation-learning method for 12-lead ECG signals, designed to reduce reliance on large labeled datasets for downstream ECG classification. TSSL exploits two structural properties of ECG data: temporally, it encourages stable representations for the same individual across time while keeping different leads distinguishable; spatially, it enforces consistency in the relationships between signals and their representations across the different leads of a single recording. Evaluated on three public ECG datasets (CPSC2018, Chapman, PTB-XL), TSSL-pretrained models approached the performance of fully supervised training while using only about 10% of the labeled data.

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.

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.

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.

Self-supervised model that performs single-frame digital-subtraction-angiography-style vessel/background separation directly from a single live (non-subtracted) coronary angiogram frame, then supports fine-tuned coronary vessel segmentation. A U-Net-style network is pretrained via an image-to-image translation objective on 58,128 unannotated angiography DICOM series (3,756 patients), then fine-tuned for vessel segmentation on just 40 expert-annotated frames, reaching a Dice of 0.828 on the held-out fine-tuning set and a new state-of-the-art Dice of 0.755 on the public XCAD benchmark. Intended to help clinicians visualize potential stenosis sites without requiring true two-frame digital subtraction acquisition.

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.

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.

LoRA-adapted domain-specialized cardiology text embedding model built on BioLinkBERT (340M parameters), identified as the top performer among 10 encoder- and decoder-style transformer architectures benchmarked head-to-head for cardiology semantic retrieval. LoRA fine-tuning on ~150,000 cardiology-textbook-derived sentence pairs raised its cardiology semantic-separation score from 0.033 (zero-shot) to 0.510, the highest of any evaluated architecture (including decoder models up to 10x larger), while remaining Pareto-optimal for the separation/throughput trade-off at 143.5 embeddings/sec and a 1.51GB memory footprint.

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).

Multimodal cardiac-sensing foundation model pretrained with generative masked pretraining on ECG, PPG, and paired clinical/machine-generated text reports from roughly 1.7 million individuals across three large-scale critical-care and outpatient ECG datasets. A channel-embedding scheme lets the same model accept any combination of 12-lead ECG, single-lead/wearable ECG, and PPG. The resulting embeddings transfer to diagnostic classification, demographic recognition, vital-sign measurement, clinical-outcome prediction, and ECG question answering. Pretrained weights require a signed academic-access agreement rather than an open download.

Generalist reconstruction foundation model for accelerating cardiac MRI (CMR) acquisition, designed to recover diagnostic-quality images from highly undersampled (8x-24x) multi-coil k-space data across heterogeneous scanners, field strengths, and cardiovascular diseases. Combines a CLIP-ViT-based module for semantic/contextual understanding of the anatomy being imaged with a physics-informed data-consistency reconstruction network, trained on MMCMR-427K -- the largest public multimodal CMR k-space database to date. Intended as an upstream substrate that feeds downstream segmentation, phenotyping, and diagnosis models (e.g. automated cardiac-phenotype extraction via nnU-Net) rather than replacing them. Released by the CMRxRecon-challenge consortium; code and the underlying database are public for academic, non-commercial use, but no separately downloadable pretrained checkpoint is provided.

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.

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.

Systematic study of domain specialization for large language models in electrocardiography, comparing supervised fine-tuning (QLoRA) against retrieval-augmented generation (RAG) as two paths to inject ECG/cardiology knowledge into open-weight Llama 3.1 models (8B and 70B). Question-answer and multiple-choice pairs were generated from ECG/cardiology literature and used both for fine-tuning and for a multi-layered evaluation (multiple-choice accuracy, text-similarity metrics, LLM-as-a-judge, and blinded human-cardiologist review). The fine-tuned Llama 3.1 70B ranked first overall, exceeding the RAG variants and Claude Sonnet 3.7 on in-distribution multiple-choice and text-similarity metrics, though RAG and Claude generalized better to semantically complex, out-of-distribution questions. Developed by AI4Health at the University of Oldenburg with Charite Berlin; the finetuning/RAG/evaluation code is public, but per the paper's data-availability statement neither the training corpus nor the fine-tuned weights are released (German copyright law, section 60d UrhG).

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).

LoRA-adapted domain-specialized cardiology text embedding model built on MPNet-base (109M parameters), identified as Pareto-optimal for balanced accuracy/throughput deployment among 10 encoder- and decoder-style architectures benchmarked for cardiology semantic retrieval. LoRA fine-tuning on ~150,000 cardiology-textbook-derived sentence pairs raised its cardiology semantic-separation score from 0.175 (zero-shot) to 0.386, while delivering 228.8 embeddings/sec at a sub-1GB (0.73GB) memory footprint, making it suitable for consumer-GPU and general-purpose medical NLP deployment where full BioLinkBERT-level accuracy is not required.

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.

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.

Multi-agent LLM framework, deployed as a Software-as-a-Medical-Device on AWS, that assists cardiologists reading 24-hour Holter/patch ECG monitoring studies. Three fine-tuned LLM agents divide the diagnostic workflow the way a cardiologist would: a table-to-text agent (Llama-3.1-8B) extracts findings from tabular arrhythmia metrics, an image-to-text agent (LLaVA-v1.5-13B) extracts findings from ECG tracing images, and a findings-to-interpretation agent (Llama-3.1-8B) synthesizes both against clinical guidelines with a fact-checking step. Each agent is instruction-tuned on cardiologist-adjudicated reports from 2,000+ real-world patients and further steered at inference with in-context demonstrations matched to the patient's age, sex and arrhythmia class. In blinded cardiologist ratings across eight clinical/security metrics (1-5 scale), ZODIAC outperformed GPT-4o, Gemini-Pro, Llama-3.1-405B, Mixtral-8x22B, and medical-specialist LLMs (BioGPT, Meditron, Med42) on every metric while using under 30B total parameters, and has been integrated into commercial ECG monitoring devices. This is a proprietary product; no public code or model weights have been released.

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.

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.

Single-lead ECG foundation model pretrained with clinically-guided contrastive learning: rather than relying on hand-labeled tasks, it uses routinely collected clinical metadata and risk scores from 161,000 MIMIC-IV-ECG patients as the training signal. Released in three sizes - Small (~448K parameters), Medium (30.7M), and Large (~296M) - and benchmarked against other ECG foundation models like ECGFounder across 18 tasks and 7 held-out datasets. Developed by Nokia Bell Labs.

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.

Domain-specialized text embedding model for clinical cardiology, built by fine-tuning the Qwen3-Embedding-8B language model with LoRA adapters via contrastive learning on cardiology textbook sentences. Reaches 99.60% top-1 accuracy on cardiology-specific semantic retrieval, nearly 16 points above the prior MedTE baseline. The training corpus draws on roughly 150,000 sentences from seven copyrighted textbooks and is not public, though the resulting model weights are freely downloadable.

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).

Reconstructs digital 12-lead ECG waveforms from scanned or photographed paper printouts, using an nnU-Net image segmentation model to trace the signal pixels followed by a Hough-transform-based reconstruction pipeline. This is a digitization tool rather than a diagnostic model - it recovers a usable signal from a paper record rather than producing a diagnosis. Won the PhysioNet/Computing in Cardiology Challenge 2024; developed by a team at the University of Oxford.

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.

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.

Vision-language foundation model fine-tuned from CLIP on more than one million private echocardiogram video-report pairs, enabling zero-shot cardiac function assessment, device identification, and image/text retrieval without task-specific training. Combines a ConvNeXt-Base video encoder with a GPT-2-style text encoder under contrastive pretraining. Training data is private, but model weights and code are public. Developed by Cedars-Sinai's Ouyang lab.

General-purpose vision foundation model for echocardiography, pretrained with a masked autoencoder combined with a periodic contrastive loss designed around the cyclical nature of cardiac motion. Validated on chamber segmentation, view classification, and disease detection, with its largest advantage over non-pretrained baselines and natural-image models like SAM appearing in low-label settings. Pretrained on roughly 290,000 echo clips from a mix of internal and public sources. Developed by Massachusetts General Hospital and Harvard Medical School.

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.

Vision-language foundation model that interprets an entire transthoracic echocardiogram study rather than a single view or video: it classifies the view type of every clip, applies view-informed attention across the full study, and generates or retrieves comprehensive study-level interpretations in English or Italian. Pretrained on a private Cedars-Sinai corpus of 12 million echo video-report pairs. Developed by the Smidt Heart Institute and Stanford's Ouyang lab.

Distills knowledge from EchoCLIP, a vision-language echocardiography model, into ECG embeddings, aiming to improve how well ECG signals alone can predict echo-derived measures of cardiac function. Combines a 1D ECG encoder with a BioBERT text encoder under a probabilistic cross-modal embedding objective that captures uncertainty. Published at MICCAI 2025 by the University of Toronto's McIntosh Lab.

GPT-style decoder-only Transformer pretrained via next-token prediction on tokenized single-lead ECG time series, producing an interpretable general-purpose model that can be fine-tuned for tasks like arrhythmia screening and beat detection. Individual attention heads are shown to respond to physiologically meaningful features such as the P-wave, and token embeddings cluster by position in the cardiac cycle. A companion PPG-pretrained model (PPG-PT) is released in the same repository. Developed at Imperial College London.

GPT-style decoder-only Transformer pretrained via next-token prediction on tokenized PPG time series, the companion model to ECG-PT (HeartGPT) in the same repository. Individual attention heads respond to physiologically meaningful waveform features such as the dicrotic notch, and the model can be fine-tuned for wearable-based cardiac screening tasks. Developed at Imperial College London.

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.

Self-supervised ECG representation learned purely from patient identity: the model is trained so that ECGs from the same patient, recorded at different times, map to nearby points in latent space, with no other labels required. Linear models trained on these representations showed a 51% average performance gain over training from scratch across sex classification, age regression, LVH detection, and AF detection. Developed by the Broad Institute's ML4H group and trained on 3.2 million private ECGs from Massachusetts General Hospital; 12-lead, lead-I-only, and lead-II-only checkpoints are all released.

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.

One of the first open foundation models for PPG signals, pretrained on over 57,000 hours (20 million segments) of publicly available data using a morphology-aware self-supervised objective. Evaluated across 20 tasks from 10 datasets spanning cardiovascular health, sleep disorders, pregnancy monitoring, and general wellbeing. Developed by Nokia Bell Labs and published at ICLR 2025.

Open-source PPG foundation model pretrained directly on real-world, field-collected wearable data rather than clean clinical signals alone, aiming for better generalization to the noise of free-living conditions. Uses a ResNet-based encoder trained with a relative contrastive (RelCon) self-supervised objective, and is directly benchmarked against PaPaGei. Developed at the University of Illinois Urbana-Champaign and published at UbiComp 2025.

Generative adversarial network that synthesizes realistic 10-second, 12-lead normal-sinus-rhythm ECGs from scratch, without using any real patient data at inference time, enabling privacy-preserving data sharing and augmentation. Uses a U-Net-style 1D deconvolutional generator with a WaveGAN-inspired discriminator. Outperformed a WaveGAN* baseline on the fraction of generated tracings classified as normal sinus rhythm by a commercial ECG interpretation algorithm. Developed by SimulaMet and Oslo Metropolitan University.

Diffusion-based generative model that synthesizes 12-lead ECGs conditioned on any of 71 PTB-XL diagnostic labels, combining a denoising diffusion process with a structured state-space (S4) sequence backbone. Outperformed GAN-based baselines (WaveGAN*, Pulse2Pulse) on both classifier-based fidelity metrics and a clinical Turing test. Developed at the University of Oldenburg.

Self-supervised ECG foundation model that adapts to varying lead combinations by patchifying 12-lead recordings across both space (leads) and time, then pretraining a ViT-B/75 encoder-decoder with a masked-autoencoder objective. Published at ICLR 2024 by VUNO Inc., and pretrained on the Chapman-Shaoxing-Ningbo dataset along with several other public 12-lead sources.