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.

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.

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.

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).

Largest federated cardiac CT analysis to date (n=8,104 scans) across a real-world federation of German university hospitals, addressing partially-labeled data across sites via a two-step semi-supervised knowledge-distillation strategy: task-specific CNNs first predict on unlabeled data per label type, then a SWIN-UNETR transformer learns from these predictions with label-specific heads. Learns a single federated model that simultaneously predicts TAVI-relevant landmarks (aortic hinge points, coronary ostia, membranous septum) and calcification from cardiac CT, improving generalizability over UNet-based baselines on downstream tasks.

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.

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.

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).

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.

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.

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.

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.

BEiT-base Vision Transformer that embeds images of standard 12-lead ECG printouts into a representation space, enabling zero-shot screening for structural heart disease by comparing a new ECG against reference case/control embedding centroids rather than requiring task-specific training. Trained on private Yale New Haven Health System ECG images and validated against the public EchoNext dataset. Aimed at scanned or legacy ECG images still common in EHR systems. Developed by Yale's CarDS Lab.