This organization contains GitHub Repositories for the Medical Event Data Standard (MEDS), a simple dataset schema for machine learning over electronic health record (EHR) data. Unlike existing tools, pipelines, or common data models, MEDS is a minimal standard designed for maximum interoperability across datasets, existing tools, and model architectures. By providing a simple standardization layer between datasets and model-specific code, MEDS can help make machine learning research for EHR data dramatically more reproducible, robust, computationally performant, and collaborative. Alongside this report, we also release several existing integrations with models, datasets, and tools, and will work actively with the community going forward for further adoption and use. See our published work in NEJM AI for more details, and please leave comments or questions via GitHub issues to help us improve this effort! Find the Contribution guidelines here. The website also contains more information.
| Project | Type | Documentation URL | Repository URL | Paper URL | PyPI | License | Description |
|---|---|---|---|---|---|---|---|
| Core MEDS | Core | Website | GitHub | OpenReview | meds | Apache-2.0 | A data standard and community for building and sharing EHR machine learning tools |
| MEDS-Reader | Preprocessing | Docs | GitHub | arXiv | meds-reader | Apache-2.0 | An optimized Python package for efficient EHR data processing achieving 10-100x improvements in memory, speed, and disk usage |
| MEDS-Transforms | Preprocessing | Docs | GitHub | Zenodo | MEDS-transforms | MIT | A set of functions and scripts for extraction to and transformation/pre-processing of MEDS-formatted data. |
| MEDS-Extract | ETL tool | Docs | GitHub | Zenodo | MEDS-extract | MIT | Config-driven ETL helpers built on MEDS-Transforms; the basis of the current dataset ETLs (v0.7 config-only layout). |
| ETL_MEDS_Template | ETL tool | GitHub | MIT | Template repository for a MEDS-Transforms/MEDS-Extract powered extraction pipeline for a custom dataset. | |||
| MESSY_synth | ETL tool | GitHub | MIT | Generates structurally faithful synthetic raw data from any MEDS-Extract v0.7 config, for ETL demos and smoke tests. | |||
| MEDS-ETL | ETL tool | GitHub | meds_etl | Apache-2.0 | Efficient ETL that supports OMOP, MIMIC, eICU, PyHealth; optional C++ backend. | ||
| OMOP_MEDS | ETL tool | GitHub | Zenodo | OMOP-MEDS | MIT | Dedicated OMOP CDM to MEDS pipeline built on MEDS-Transforms. | |
| MEDS on FHIR (fhir2meds) | ETL tool | Docs | GitHub | Zenodo | fhir2meds | none set | Converts HL7 FHIR (v4/v5) resources into sharded, schema-validated MEDS Parquet. |
| EHR2MEDS | ETL tool | GitHub | MIT | Formats raw EHR dumps and converts them to MEDS (from the BONSAI / CORE-BEHRT group). | |||
| EHR2Trace | ETL tool | Website | GitHub | arXiv | ehr2trace | Apache-2.0 | Auditable, source-linked conversion of hospital EHR data for patient world models and clinical agents. |
| meds_subsetter | Preprocessing | GitHub | MIT | Nested, hash-based subject subsetting, resharding, sizing and fingerprinting of MEDS datasets (marked AI-generated). | |||
| ACES | Task extraction | Docs | GitHub | ICLR 2025 / arXiv | es-aces | MIT | A package and configuration language for reproducible extraction of task cohorts for machine learning over event-stream datasets |
| meds-random-task-sampler | Task extraction | GitHub | meds-random-task-sampler | MIT | Model-independent, reproducible generation of code-by-horizon query tasks from MEDS datasets. | ||
| ehrlint | Task extraction | GitHub | Apache-2.0 | Audits a finished MEDS or OMOP prediction task for label leakage before training. | |||
| MEDS-Tab | Modeling | Docs | GitHub | arXiv | meds-tab | MIT | A library designed for automated tabularization, data preparation with aggregations and time windowing. |
| MEDS-Torch | Modeling | Docs | GitHub | OpenReview | meds-torch | MIT | Advancing healthcare machine learning through flexible, robust, and scalable sequence modeling tools. |
| meds-torch-data | Modeling | Docs | GitHub | Zenodo | meds-torch-data | MIT | PyTorch dataset objects for sequence modeling over MEDS datasets (used by MEDS-EIC-AR). |
| FEMR | Modeling | GitHub | MOTOR (arXiv) | femr | Apache-2.0 | A Python package for manipulating longitudinal EHR data for machine learning, with a focus on supporting the creation of foundation models and verifying their presumed benefits in healthcare. Reads MEDS v0.3. | |
| meds_to_cehr | Modeling | GitHub | MIT | Converts MEDS datasets into CEHR-BERT / CEHR-GPT patient sequences. | |||
| meds2text | Modeling | GitHub | arXiv | Apache-2.0 | Transforms MEDS data into textual representations for LLMs. | ||
| TwinWeaver | Modeling | Docs | GitHub | arXiv | twinweaver | Apache-2.0 | A package for training & evaluating LLMs on longitudinal clinical data. |
| WeavEHR | ETL tool | Wiki | GitHub | weavehr | MIT | Declarative YAML framework (Polars-based successor to ricu) that extracts and harmonises ICU data into MEDS concepts across MIMIC-IV, eICU, NWICU, HiRID, SICdb and AmsterdamUMCdb (via OMOP). |
|
| MEDS-Evaluation | Evaluation | Website | GitHub | meds-evaluation | MIT | Evaluation pipeline for MEDS model predictions in a shared pyarrow schema. |
|
| MEDS-DEV | Benchmark | Website | GitHub | MEDS-DEV | MIT | Decentralized, extensible validation benchmark for ML models on MEDS-formatted data. | |
| MEDS-Inspect | Exploration | GitHub | Zenodo | MEDS-Inspect | MIT | A package to interactively inspect your MEDS data. | |
| suMEDS | Exploration | Docs | GitHub | Zenodo | suMEDS | MIT | Summarizes a MEDS dataset into a compact, digestible code catalog. |
| meds_summary_stats | Exploration | GitHub | meds_summary_stats | MIT | Fast, privacy-safe summary statistics over MEDS datasets for validating and regression-testing ETLs. | ||
| MEDS_visualizations | Exploration | GitHub | Zenodo | MIT | Visualization tools for MEDS datasets. | ||
| MEDS Testing Helpers | Testing | Docs | GitHub | meds_testing_helpers | MIT | Testing, benchmarking, and synthetic data generation helpers for MEDS tools, pipelines, and models. | |
| MEDS Ontology | Semantic | Website | GitHub | Zenodo | MIT | OWL ontology encoding the MEDS conceptual data model. | |
| meds2rdf | Semantic | GitHub | Zenodo | MIT | Converts MEDS datasets into RDF using the MEDS Ontology. |
| Model | Type | Params | Pretraining Objective | Input / Tokenization | Zero-shot / Generative | Originally Published | Code License | Pretrained Weights | Paper | Code | MEDS Support |
|---|---|---|---|---|---|---|---|---|---|---|---|
| MEDS-TAB | Tabularization + XGBoost AutoML (also LR, RF, kNN) | n/a | None (supervised only) | Static features + rolling-window aggregates (count, sum, min, max, …) | No | arXiv 2024 | MIT | n/a | arXiv 2411.00200 | GitHub | Native; in MEDS-DEV |
| GenHPF | Hierarchical transformer (event encoder + aggregator) | Not stated | Optional SSL (SimCLR, Wav2Vec 2.0, MLM, SpanMLM); MEDS-DEV entry is supervised only | Code descriptions as text (Bio_ClinicalBERT tokenizer), digit-wise numbers, time intervals | No | IEEE JBHI 2023 (arXiv 2022) | MIT | None | arXiv 2207.09858 | GitHub | Converter (genhpf-preprocess-meds); in MEDS-DEV |
| CEHR-BERT | BERT encoder | Not stated (5 layers, 128-d) | MLM + visit-type prediction | Concept codes + artificial time tokens, visit tokens, age/time embeddings; optional values | No | ML4H 2021 | MIT | None | PMLR | GitHub | Converter via meds_reader; in MEDS-DEV |
| CEHR-GPT / CEHR-XGPT | GPT decoder | Not stated (XGPT: 16 layers, 768-d, ctx 4096) | Next-token; XGPT adds time-decomposition + time-to-event | CEHR-BERT-style sequence with demographic prompt, artificial time tokens, value tokens | Yes (zero-shot, synthetic trajectories) | arXiv 2024 (XGPT: arXiv 2025) | MIT | None | arXiv 2402.04400, XGPT | GitHub | Converter via meds_reader; MEDS-DEV PRs closed (dataset-specific conversion) |
| MOTOR | Transformer + piecewise-exponential TTE head | 143M | Time-to-event over 8,192 code tasks | femr hierarchical code tokenizer, quantile-binned values, rotary age embedding | No | ICLR 2024 | Apache-2.0 (femr) | Gated, CC-BY-NC 4.0 | arXiv 2301.03150 | GitHub | Native via femr (MEDS v0.3 only); unmerged MEDS-DEV branch |
| CLMBR-T-base | Transformer decoder | 141M | Next-code prediction | 65,536-token OMOP code vocabulary | Not evaluated | NeurIPS 2023 D&B | Apache-2.0 | Gated, CC-BY-NC 4.0 | arXiv 2307.02028 | GitHub | Native via femr 0.2.3; not in MEDS-DEV |
| Context Clues | GPT / Llama / Mamba / Hyena | ~116–127M | Next-token prediction | CLMBR tokenizer (39,811 tokens: codes, categorical values, numeric deciles) | Not evaluated | ICLR 2025 | Apache-2.0 | Gated, CC-BY-NC 4.0 (16 checkpoints) | arXiv 2412.16178 | GitHub | MEDS via meds_reader → hf_ehr; not in MEDS-DEV |
| ETHOS / ETHOS-ARES | GPT-2-style decoder | ~45M (ARES) | Next-token prediction | Hierarchical codes (ICD-10, ATC), 10 quantile value tokens, 19 time-interval tokens | Yes (zero-shot by trajectory rollout, synthetic data) | npj Digit Med 2024; ARES: GigaScience 2025 | MIT | None (MIMIC restrictions) | ETHOS, ARES | GitHub | Native MEDS; MEDS-DEV issue open |
| MEDS-EIC-AR | Llama decoder ("everything-is-code") | Presets ~4M–400M (estimate) | Next-token prediction | MEDS codes, decile-binned values, time-delta tokens | Yes (trajectory generation) | Software only (Zenodo 2025) | MIT | None | — | GitHub | Native MEDS; MEDS-DEV PR #313 open |
| EveryQuery | Llama-style decoder | ~22M | Task-conditioned query pretraining ("will code c occur in window?") | MEDS codes with binned values; time as rotary positions | Yes (promptable zero-shot classification + TTE, one forward pass) | arXiv 2026; ICML 2026 workshop | MIT | None | arXiv 2603.07900 | GitHub | Native MEDS; not in MEDS-DEV |
| MEDS-Torch | Framework (transformer encoder/decoder, LSTM, …) | Configurable | OCP, EBCL (contrastive), STraTS forecasting, autoregressive, or supervised | Triplet, Text-Code, or Everything-In-Code tokenizers | Autoregressive variants | NeurIPS 2024 workshop (TSALM) | MIT | None | OpenReview | GitHub | Native MEDS; not in MEDS-DEV |
| BONSAI (successor of CORE-BEHRT) | BERT encoder (RoPE, SwiGLU, Time2Vec) | Not stated (CORE-BEHRT: 6 layers, 192-d) | MLM (BONSAI also next-token) | ICD-10 / ATC codes, age, timestamps; BONSAI can bin values | No | CORE-BEHRT: ML4H 2024; BONSAI: JOSS 2025 | MIT (BONSAI); none (CORE-BEHRT) | None | arXiv 2404.15201 | GitHub | BONSAI native MEDS; original CORE-BEHRT is not; not in MEDS-DEV |
| Model | Parameters | Architecture | Pretraining Objective | Context Length | Pretraining Data | Released | License | Access | Weights | Paper | Contact |
|---|---|---|---|---|---|---|---|---|---|---|---|
| CLMBR-T-base | 141M | Transformer decoder (12 layers, 768 hidden) | Autoregressive next-code prediction | 496 | 2.57M patients, Stanford Medicine (STARR, OMOP-CDM, structured only) | Dec 2023 | CC-BY-NC 4.0 | Gated: CITI "Data or Specimens Only" cert + EHRSHOT Credentialed Health Data License | HF | EHRSHOT, NeurIPS 2023 | Jason Fries (jfries@stanford.edu, model card) |
| MOTOR-T-base | 143M | Transformer (12 layers, 768 hidden) | Time-to-event (piecewise exponential over future codes) | 496 | 2.57M patients, Stanford Medicine (STARR, OMOP-CDM) | Feb 2024 | CC-BY-NC 4.0 | Gated: CITI cert + EHRSHOT license | HF | MOTOR, ICLR 2024 | Jason Fries (jfries@stanford.edu, model card); Ethan Steinberg |
| Context Clues: GPT | ~116M | GPT decoder | Autoregressive next-code prediction | 512, 1,024, 2,048, 4,096 | 2.57M patients / ~2B tokens, Stanford Medicine (STARR) | Feb 2025 | CC-BY-NC 4.0 | Gated: CITI cert + EHRSHOT license | HF collection | Context Clues, ICLR 2025 | Michael Wornow; Nigam Shah (nigam@stanford.edu, EHRSHOT paper) |
| Context Clues: Llama | ~123M | Llama decoder | Autoregressive next-code prediction | 512, 1,024, 2,048, 4,096 | same as above | Feb 2025 | CC-BY-NC 4.0 | Gated: CITI cert + EHRSHOT license | HF collection | Context Clues, ICLR 2025 | same as above |
| Context Clues: Mamba | ~121M | Mamba (state-space model) | Autoregressive next-code prediction | 1,024, 4,096, 8,192, 16,384 | same as above | Feb 2025 | CC-BY-NC 4.0 | Gated: CITI cert + EHRSHOT license | HF collection | Context Clues, ICLR 2025 | same as above |
| Context Clues: Hyena | ~125M | Hyena (long convolution) | Autoregressive next-code prediction | 1,024, 4,096, 8,192, 16,384 | same as above | Feb 2025 | CC-BY-NC 4.0 | Gated: CITI cert + EHRSHOT license | HF collection | Context Clues, ICLR 2025 | same as above |
We now have an official MEDS MIMIC-IV demo dataset on Physionet, which allows you to familiarize yourself with MEDS without running or downloading any additional software.
| Dataset | Stays | Version | Frequency | Origin | Originally Published | License | Repository Link | MEDS ETL | Full Dataset Name |
|---|---|---|---|---|---|---|---|---|---|
| AUMCdb | 23,000 | v1.0.2 | up to 1 minute | Netherlands | 2019 | AmsterdamUMCdb End User License (non-commercial) | DANS | Github | Amsterdam University Medical Center Database |
| eICU | 201,000 | v2.0 | 5 minutes | USA | 2017 | PhysioNet Credentialed | PhysioNet | Github | eICU Collaborative Research Database |
| HiRID | 34,000 | v1.1.1 | 2 / 5 minutes | Switzerland | 2020 | PhysioNet Contributor Review | PhysioNet | Github | High-Resolution ICU Dataset |
| INSPIRE | 130,000 | v1.2 (ETL now targets v1.4.2) | 5 minutes (OR); ~hourly (ICU) | South Korea | 2024 | Korea Credentialed Health Data License | PhysioNet | Github | INformative Surgical Patient dataset for Innovative Research Environment |
| MIMIC-IV | 73,000 | v3.1 | ~1 hour | USA | 2020 | PhysioNet Credentialed | PhysioNet | Github | Medical Information Mart for Intensive Care IV |
| NWICU | 25,000 | v0.1.0 | Not specified | USA | 2024 | PhysioNet Credentialed | PhysioNet | Github | Northwestern ICU Database |
| SICdb | 27,350 | v1.0.8 | 1 minute | Austria | 2023 | PhysioNet Contributor Review | PhysioNet | Github | Salzburg Intensive Care Database |
| CDSL | 4,479 (hospitalizations, not ICU-only) | v1.0.0 | Not specified (sparse, ~50 vitals rows per episode) | Spain | 2024 | PhysioNet Contributor Review | PhysioNet | Github | COVID Data for Shared Learning |
| ALOTT | 15,000+ (hospital admissions, 270 telemetry beds) | v1.0.0 | 2 seconds (vitals); 60–240 Hz (waveforms, not extracted by ETL) | USA | 2025 | PhysioNet Restricted | PhysioNet | Github | ALarms, Outcomes Telemetry with Timing |
| PIC | 13,941 (ICU stays; paediatric) | v1.1.0 | ~daily (charted); 5 minutes (intra-op) | China | 2019 | PhysioNet Credentialed | PhysioNet | Github | Paediatric Intensive Care Database |
| Zigong | 2,790 (ICU patients with infection) | v1.1 | Not specified | China | 2021 | PhysioNet Credentialed | PhysioNet | Github | Zigong Fourth People's Hospital ICU Infection Database |
| EHRSHOT | 6,739 patients (921,499 visits; longitudinal, not ICU-only) | Original / MEDS release (EHRSHOT_MEDS.zip) |
Not specified (coded structured events) | USA | 2023 | EHRSHOT Credentialed Health Data License (modeled on PhysioNet 1.5.0) | Redivis | Native MEDS release (converter) | EHRSHOT: An EHR Benchmark for Few-Shot Evaluation of Foundation Models |
Moreover, we have these ETLs for other formats:
- OMOP ETL: https://github.com/rvandewater/OMOP_MEDS
- FHIR ETL: https://github.com/rvandewater/FHIR_MEDS
If you use MEDS in your work, please cite it like this:
@article{mcdermottMEDSEmergingData2026,
title = {{{MEDS}} --- {{An Emerging Data Standard}} and {{Ecosystem}} for {{Health AI Research}}},
author = {McDermott, Matthew B. A. and Steinberg, Ethan and Fries, Jason A. and {van de Water}, Robin P. and Pang, Chao and Rockenschaub, Patrick and Renc, Pawel and Oh, Jungwoo and Stankevi{\v c}i{\=u}t{\.e}, Kamil{\.e} and Xu, Justin and Pollard, Tom J. and Oufattole, Nassim and Wornow, Michael and Bergamaschi, Teya S. and Jeong, Hyewon and Lee, Simon A. and Jeanselme, Vincent and Klein, Kiril V. and Odgaard, Mikkel and Montgomery, Maria E. and Sitek, Arkadiusz and Nielsen, Mads and Chiang, Jeffrey N. and Dagan, Noa and Kohane, Isaac and Joshi, Shalmali and Choi, Edward and Shah, Nigam H.},
year = 2026,
month = may,
journal = {NEJM AI},
volume = {3},
number = {6},
pages = {AIra2501253},
publisher = {Massachusetts Medical Society},
doi = {10.1056/AIra2501253},
urldate = {2026-05-28}
}Tools that are planned to be compatible with MEDS:
