OneEHR
EHR AI platform MedX-PKU/OneEHR

Scope

Use this page to understand what OneEHR covers and where to connect external tools.

What OneEHR Covers

Area Included
Input data Standard CSV tables for dynamic events, static covariates, and labels
Preprocessing Event binning, feature encoding, train-split-fitted preprocessing steps, and patient-level splits
Models 42 built-in model config names across tabular ML, DL, irregular-time, multimodal, KG-enhanced, and survival families
Systems LLM and agent system definitions that write predictions into the same test artifact
Evaluation Metrics, bootstrap confidence intervals, calibration, fairness, feature importance, statistical tests, and missing-data summaries
Outputs Run manifests, Parquet tables, JSON reports, checkpoints, and figures
Dataset conversion MIMIC-III, MIMIC-IV, and eICU converters for supported tasks
Medical codes ICD-9/10 parsing, CCS grouping, ATC hierarchy, and code mapping helpers

Boundaries

OneEHR is an experimentation package. It does not provide clinical decision support, protected health information hosting, credential management, or access to licensed datasets.

Raw clinical datasets must be obtained and stored according to their own data-use agreements. OneEHR converters read local files and write derived CSV tables.

Built-in model baselines train from the run config and saved preprocessing artifacts. Models that use optional external assets, such as external KG files or text embeddings, record those paths through explicit model parameters and checkpoint metadata.

Public Terms

Use these terms consistently across docs, examples, and tutorials:

Prefer concrete descriptions of what a user can run or inspect. Avoid competitive claims and broad statements that do not change how someone uses the package.