Add diagnostic-delay concept module (cohort + delay measures + data-quality report)

[developer-rpai]

What this is

A new concept module, mimic-iv/concepts/diagnosticdelay/, for reproducible cohort extraction and diagnostic-delay measurement on MIMIC-IV:

• suspected_sepsis_cohort.sql — one row per ICU stay: adults admitted via the emergency department with at least one suspected-infection episode (antibiotic + microbiology culture pair, per suspicion_of_infection), with anchor timestamps and demographics.
• diagnostic_delay.sql — delay in hours between hospital/ICU admission and first infection recognition (suspected-infection time, first antibiotic, first culture); an onset-window classification (pre_admission / present_on_admission / hospital_onset); and temporal-plausibility flags (1 = implausible, NULL = unassessable) for out-of-order or missing timestamps.
• data_quality_report.py — generates a Markdown/JSON data-quality report (cohort coverage, missingness, flag prevalence, delay distributions with median/p25/p75) from the two tables; runs on DuckDB or PostgreSQL with portable SQL.
• README.md + concept-index entry — documents the cohort definition, measures, validation, and limitations.

Why

ICD codes in MIMIC carry no timestamp, so "when was this diagnosed relative to admission?" is a recurring community question (e.g. #1843 on diagnosis timing for sepsis/AKI research). This module answers a tractable version of it by proxying diagnosis time with the clinically observable recognition events already used in the Sepsis-3 definition. The plausibility flags address the class of problem raised in #2168 (implausible values invisible to anyone who doesn't check before aggregating): flagged stays can be investigated rather than silently averaged in.

Validation (what ran, what didn't)

I do not have credentialed MIMIC-IV access, so validation was done without real data — an end-to-end run on MIMIC-IV (v3.1) is still outstanding:

1. sqlglot parse check of both queries in the BigQuery dialect (.github/scripts/check_sql_syntax.py) — pass.
2. sqlfluff lint with the repo's .sqlfluff config (sqlfluff 4.1.0, same as CI) — pass, no violations.
3. Transpiled both queries BigQuery → DuckDB and BigQuery → PostgreSQL with the repo's mimic_utils transpiler; both parse in the target dialects.
4. Executed the DuckDB build against hand-built synthetic fixtures mirroring the MIMIC-IV schema (patients, admissions, icustays, derived suspicion_of_infection): 16 assertions covering a standard present-on-admission case, a hospital-onset case, a pre-admission-suspicion case (flagged), multiple episodes per stay, pediatric exclusion, non-ED admission exclusion, a stay with no suspected infection (excluded), and a NULL icu_outtime case (flag unassessable) — all pass, delays match hand computation.
5. Ran data_quality_report.py against the fixture database (with a raw-schema coverage denominator): Markdown/JSON outputs verified against hand-computed values (4/5 = 80% capture, flag counts, medians/p25/p75).

Notes for reviewers

• Only mimic-iv/concepts/ sources are touched. The transpiled concepts_postgres/concepts_duckdb outputs are intentionally not included — the regenerate-dialects bot produces those on merge. The duckdb.sql/postgres-make-concepts.sql build lists are likewise left for maintainers to extend.
• The module depends only on mimiciv_hosp + mimiciv_icu + derived concepts (no MIMIC-IV-ED module), so it runs on the standard build. ED arrival time is deliberately not used; the README notes this as a limitation.
• Draft PR — happy to adjust naming, column choices, or the onset-window thresholds to fit maintainer conventions.