De-identified electronic health records (EHR) from critical care patients (ICU stays at Beth Israel Deaconess Medical Center). Includes demographics, vital signs, lab results, medications, etc., for tens of thousands of ICU admissions. Access: Free for research with data use agreement and training. MIMIC-IV (2008–2019) is the updated version of MIMIC-III.

Multi-center intensive care unit database from 208 hospitals across the US. Contains patient physiology, treatments, and outcomes. Available via PhysioNet with similar access requirements.

A large public imaging dataset of 112,000+ frontal chest X-rays from the NIH Clinical Center, annotated with 14 common thoracic pathology labels. Widely used for training and benchmarking medical image models.

A collection of 224,316 chest radiographs from 65,240 patients, labeled for 14 observations via automated report analysis. A subset with radiology reports paired to images (CheXpert Plus) is also available. Access: Public for research use (hosted by Stanford AIMI).

An NIH-funded repository of medical imaging datasets, primarily cancer-related (CT, MRI, digital pathology, etc.). Many collections include expert annotations or lesion segmentations. Access: Open to public.

A landmark cancer genomics program that molecularly characterized ~11,000 tumors across 33 cancer types. Generated 2.5 petabytes of genomic and clinical data, publicly available for research. Access: Through the NCI Genomic Data Commons.

PhysioNet hosts a variety of physiological waveform datasets: ECG recordings (MIT-BIH Arrhythmia Database), EEG, wearable sensor data, and more. These support research in cardiovascular and neurological signal analysis.

The i2b2 and n2c2 challenges have released de-identified clinical text corpora with annotations for NLP tasks. Access: Available to researchers via data use agreements.

A biomedical NLP transformer model pre-trained on millions of biomedical texts (PubMed articles). Can be fine-tuned for tasks like extracting clinical concepts or classifying biomedical documents.

A generative Transformer model pre-trained on 15M+ PubMed abstracts and specialized for biomedical language. Released openly on Hugging Face Hub.

An open-source convolutional neural network for COVID-19 diagnosis from chest X-rays. Released as a reference implementation for researchers to build on. For research use only.

Open-sourced protein structure prediction model. The AlphaFold Protein Structure Database provides predicted structures for over 200 million proteins freely to the public, enabling drug target identification and disease mutation analysis.

A library with pre-trained models for chest X-ray analysis trained on large public datasets. Open-source (PyTorch-based), allowing researchers to apply deep learning to imaging data without training from scratch.

The nnU-Net framework with pre-trained weights for medical segmentation tasks. Includes open models for automated segmentation of lung tumors on PET/CT and other organs, serving as strong baselines for new imaging projects.

A PyTorch-based open-source framework designed for healthcare imaging AI. Provides domain-optimized functionality for reading medical images (DICOM, NIfTI), data augmentation, and deep learning model architectures like UNet variants.

A self-configuring open-source framework that automatically configures and trains a UNet-based neural network for any new segmentation dataset. Won numerous medical imaging challenges and greatly democratizes medical image segmentation.

A mature open-source NLP toolkit for clinical text developed by the Mayo Clinic. Ingests clinical notes and identifies medical concepts like disorders, drugs, and symptoms, mapping them to standardized vocabularies.

An extension of the spaCy NLP library with clinical-specific components including section detection, contextual negation detection (NegEx), and extraction of clinical observations. Easier to use than cTAKES for Python developers.

An open-source Python toolkit for developing predictive models on healthcare data. Implements over 30 predictive modeling algorithms under a unified API, improving reproducibility across health AI projects.

A popular open-source data annotation tool supporting labeling of text, images, and audio via a web interface. Useful for building custom medical annotation workflows including radiology bounding boxes and clinical note labeling.

The two leading open-source deep learning frameworks extensively used in medical AI research. Both are free, have Python APIs, and support GPU acceleration, lowering the barrier for implementing novel AI models on health data.

OHDSI is an open science community behind the OMOP common data model. Their tools like Atlas enable cohort identification and analysis on population-scale observational clinical data, facilitating standardized AI model development across institutions.