OneEHR: Fair and reproducible AI studies using electronic health records

What is it about?

Electronic health records capture a patient’s care over time, but small choices about who is included, what counts as an outcome, when information is used, and how patients are divided between development and testing can change which AI method appears best. OneEHR is an open-source toolkit that keeps those choices consistent and visible. With one configuration, researchers can prepare data, run conventional machine-learning, deep-learning, language-model, and AI-agent approaches, compare them on the same patients and outcomes, and inspect saved predictions, performance, calibration, subgroup results, and figures. This publication presents a three-hour hands-on tutorial that teaches the full workflow and gives participants reusable examples for their own clinical AI studies.

Featured Image

Photo by Luke Chesser on Unsplash

Photo by Luke Chesser on Unsplash

Why is it important?

Clinical AI is moving rapidly from individual prediction models to large language models and agent systems, yet comparisons can be misleading when methods use different patient groups, preprocessing, labels, data splits, or evaluation rules. OneEHR provides a shared, auditable experiment contract so that differences in results are more likely to reflect the methods themselves rather than hidden pipeline choices. It brings data preparation, testing, analysis, and reporting into one reproducible workflow, while allowing 42 built-in models and external language-model or agent systems to produce comparable outputs. This can make clinical AI studies easier to reproduce, extend, and audit—and help build more trustworthy evidence before deployment.

Perspectives