feat(meds): OMOP → MEDS conversion for ML-native EHR encoders

[bschilder]

Summary

Adds synthlab.meds, a thin wrapper over the community meds_etl package so SynthLab output (Synthea → OMOP CDM) can feed directly into MEDS-native foundation models like SMB-v1 and MOTOR.

Pipeline

SyntheaRunner.run()                   # synthlab.synthea (existing)
    → Synthea CSVs
convert_synthea_to_omop(...)          # synthlab.synthea (existing)
    → OMOP v5.4 CSVs
convert_omop_to_meds(...)             # synthlab.meds (new in this PR)
    → MEDS parquet shards
load_meds_events(meds_dir)            # synthlab.meds (new)
    → polars DataFrame ready for SMB-v1 / MOTOR / ...

Public API

• MedsConvertConfig dataclass — mirrors the meds_etl_omop CLI surface (backend, num_shards, num_proc, overwrite)
• convert_omop_to_meds(config) — invokes meds_etl_omop as a subprocess, surfaces stderr on failure, refuses to clobber existing output unless asked
• load_meds_events(meds_dir, subject_ids=None) — polars-native reader for the sharded parquet, sorted by (subject_id, time); cohort filter pushed into the scan
• get_meds_cache_dir() / get_meds_info() / print_meds_info() — matches the pattern every other synthlab submodule uses

Install

pip install synthlab[meds] (new optional extra; pulls meds_etl>=0.3 + meds>=0.4).

Infra

• New tests/ directory + test_meds.py — 15 tests covering cache-dir, info dict, config validation, all error paths (ImportError, FileNotFoundError, FileExistsError), shard reading, subject filtering, and meds_etl_omop console-script registration
• New .github/workflows/test.yml — runs pytest on every PR + push to main (previously no CI existed)
• pyproject.toml gets new [meds] and [tests] extras (rolled into [all])
• README gets a new "MEDS conversion" section with a quickstart snippet

Test plan

• pytest tests/ passes locally (15/15) on Python 3.11
• import synthlab.meds clean in fresh env
• convert_omop_to_meds raises clear errors for missing meds_etl, missing OMOP dir, and existing output dir without overwrite=True
• load_meds_events round-trips a synthetic parquet shard + respects subject_ids filter
• GHA CI green on this PR

Motivation

Enables downstream work in smb-protopheno where each cohort's synthetic EHR needs to be encoded by SMB-v1 as part of a joint genomics + EHR model. Without this, every consumer of SynthLab's output has to re-implement the OMOP → MEDS ETL.

🤖 Generated with Claude Code