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.

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.

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.

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.

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.

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.

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.

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.

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.