suMEDS

suMEDS scans a MEDS dataset and writes the code information needed to discover and define downstream tasks:

• canonical code descriptions, parent codes, modifiers, and extension metadata;
• event occurrence counts;
• unique-subject occurrence counts;
• optional counts per canonical subject split, as rows or columns beside totals;
• configurable masking or removal of rare code cells;
• optional descriptions, ancestor/descendant/sibling codes, domains, and concept identifiers from OHDSI Athena CSV files or PostgreSQL.

Data flow

metadata/dataset.json ── code modifier declaration ─┐
data/**/*.parquet ── projected lazy scan ── counts ├─ privacy policy ── Parquet
metadata/codes.parquet ── descriptions/parents ─────┤
metadata/subject_splits.parquet ── optional split ──┤ Parquet / CSV / JSON
Athena CSV/PostgreSQL ── optional released-code enrichment ──┘

Only subject_id, code, declared code modifiers, and optionally split are used. Polars performs projection pushdown and streaming Parquet writes; no patient-level result is collected by Python. Standard JSON arrays are assembled line by line, so JSON output is bounded-memory too.

Compatibility

The package uses the official meds schema classes and targets MEDS 0.4.x. The included MIMIC-IV demo identifies itself as MEDS 0.3.3 and remains compatible with the fields used by this package.

Start with Usage, then review the privacy model before sharing an output.