Healthcare & biomedical training data
Find governed, provenance-tracked datasets that match a model-development need — imaging, clinical tables, and patient event streams.
Featured datasets · 4 available to your account
De-identified ICU + hospital EHR tables (labs, admissions, chart events, prescriptions). Rich clinical schema with well-known join keys — a strong fit for clinical NLP, phenotyping, and structured-EHR model development.
- Clinical phenotyping & risk models
- Structured-EHR representation learning
- Lab-result and medication reasoning tasks
The same MIMIC-IV cohort reshaped into the MEDS event-stream standard with ready-made train/tuning/held_out splits. Ideal when you want temporal event sequences per patient rather than relational tables.
- Sequence models over patient event streams
- Foundation-model pretraining on clinical events
- Benchmarking with predefined data splits
Chest X-ray images organised into folders by NIH Finding Label (Pneumonia, Effusion, Atelectasis, Infiltration, No_Finding), each image linked to CSV metadata and a JSON sidecar. A quick starting point for chest-X-ray classification and multimodal imaging experiments.
- Chest X-ray classification (e.g. pneumonia)
- Weakly-labelled medical image pretraining
- Image↔metadata linkage and multimodal tasks
Structured target–disease–drug association knowledge as Parquet tables. Useful for grounding biomedical models, building knowledge graphs, and linking drug mechanisms to diseases. Open (CC0) and non-patient data.
- Target/disease/drug link prediction
- Biomedical entity & relation grounding
- Retrieval over drug mechanism-of-action evidence