PSDL Inspector (MIT): authoring + governance layer for OMOP phenotypes and SQL preflight — looking for feedback

Developers
1

Hi all,

I’ve been building an open-source tool and would genuinely appreciate criticism from people who know this ecosystem better than I do — including “this already exists, use X.”

The problem I keep running into is that cohort and detection logic often ends up buried in SQL or notebooks. Reviewers (and future-you) have to reverse engineer what a query actually does, and sometimes you only discover performance issues after the query has been running for a long time.

PSDL Inspector is my attempt at an authoring + governance layer above SQL.

A clinical scenario can be authored in plain English, through a guided builder, or directly in a small YAML DSL (PSDL). It produces:

  • Validation plus a visual graph / semantic outline of signals → trends → decision logic.
  • OMOP terminology anchoring. Terms are mapped to concept_ids using BioLORD-2023 embeddings and a clinical reranker, with confidence levels and explicit “unanchored” warnings.
  • A checksummed bundle containing the scenario, anchors, and audit metadata, together with an IRB-ready Word export.
  • A deterministic (zero-LLM) SQL preflight that estimates scale, lineage, and potential bottlenecks and returns GO / CAUTION / BLOCK before execution. Optionally, it can use EXPLAIN against your own database without executing the query or reading rows. I’ve been testing this on a local MIMIC-IV-in-OMOP instance.

I don’t see this as replacing ATLAS or the Phenotype Library. My mental model is that it sits upstream as an authoring and governance layer and emits a portable, anchored artifact.

I’m mainly trying to understand where this overlaps with existing OHDSI workflows, where it’s redundant, and where it might actually be useful.

Repo (90-second demo + docs):

Some questions I’d particularly appreciate feedback on:

  1. How does embedding-based terminology anchoring compare with your current concept mapping workflow?

  2. Would a checksummed, certified bundle be useful for IRB review, reproducibility, or governance, or is this already solved in your environment?

  3. Would a SQL preflight layer be valuable on your warehouse (Postgres, BigQuery, Databricks, Redshift, etc.), and what would you want it to tell you?

Thanks in advance. Happy to be told I’m reinventing something.