B2SLab — Bioinformatics and Biomedical Signals Laboratory Universitat Politècnica de Catalunya UPC IRIS

EHRMENEIA

A Benchmark for Foundation Models on Longitudinal Electronic Health Records

Leaderboard

Loading submissions…

Dimension I · Level 1

Predictive Competence

How well does the model predict the next event?

Dimension II · Levels 2–3

Clinical and Population Fidelity

How faithfully do generated futures reproduce clinical events across time?

Dimension III · Level 4

Representation Transferability

How efficiently can a model be adapted to new clinical tasks?

Models

Submission requirements

Provide the model identity, pretraining descriptors and references, evaluation cohort and protocol declarations. Use null for unpublished pretraining statistics; they appear as “not reported”. Include a public vocabulary mapping when available. Model weights and patient-level data remain local.

Submit results

Complete the metadata template before running the benchmark. The pipeline validates a compact submission with aggregate results and the public vocabulary mapping.

  1. Download the metadata template, complete it and save it as metadata.json in your submission directory.
  2. Run the reference pipeline for the levels you wish to report.
  3. Select the generated results_submission.zip (or <output_name>_submission.zip). It contains the result JSON, the declared vocabulary mapping when supplied and, when Level 3 is completed, aggregate population curves.

The selected results are inspected in your browser first. Submitting creates a separate result entry and makes its aggregate results, metadata, terminology mapping and aggregate curves publicly available. Model weights remain local. Individual Parquet records are never needed by this website.