12-1 · Crisis prevention · 2026-11-14

Suicide Prevention Research for People with Mental Illness

Cohort and simulationSynthetic checks matchedEvaluation data partly established

Evaluate event timing and references to supporting evidence in synthetic longitudinal records.

This is a research record, not a help service. If you are in Korea and need to talk to someone now, the suicide prevention counselling line 109 can be reached by phone. Elsewhere, contact your local emergency number.

Research question and target

Align event markers and note excerpts by time in offline research. Test for information leakage after the reference time and missed signals across subgroups using controlled examples.

Internal target: At least 90% recall for acute crisis signals
Programme: Crisis prevention
Evidence on this record: Synthetic harness (prevent12)Common synthetic checks

Evaluation data

MIMIC-IV-Note
2.2, 2023-01-06
Blocked
CDC NSSP annual national aggregate table
MMWR 75(24), 2021–2025
Manifest locked

Evaluation registry · Labels and split definitions (KR)

Synthetic harness · prevent12

Early detection from de-identified longitudinal records

Recall on the synthetic test set, weighted for censoring, was 0.9155 against the internal target of 0.90. The alert burden was 46.1 per 100 patient-months and episode-level precision was 0.079.

Recall against latent reference: 0.9324
Positives in the test set: 167
Expected calibration error: 0.020
Subgroup audit: GAP_EXCEEDS_TOLERANCE

Limits

Sequences are generated by code. No clinical note, real cohort or domain vocabulary is read or stored.
The evaluation watches the highest score in the window before an event; a single landmark time was not used.

Synthetic inputs only (synthetic_only). These are virtual metrics on synthetic inputs. No real cohort has been evaluated, clinical validity has not been assessed and no auditor has signed the result. A separate audit program passed 4 of 4 gates.

G1 · Frozen inputs and submission structureG2 · Schema, policy rules and transition completenessG3 · Independent recalculation of thresholds and metricsG4 · Virtual metric verdict and result hash

Audit record (JSON) · Run output (JSON) · 12-1 design specification: early detection of acute crises (KR) · 12-1 to 12-4 replay record (KR)

python3 -B -m prevent12.orchestrate --output evidence/prevent12-run.json
python3 -B -m prevent12.audit --run evidence/prevent12-run.json --output evidence/prevent12-audit.json
python3 -B -m unittest discover -s tests -p 'test_prevent12.py' -v

Sources and eligibility

PhysioNet MIMIC-IV-Note
Credentialed access
Contains deidentified discharge summaries and radiology reports. Longitudinal psychiatric notes and subsequent crisis labels require separate data.

Boundary cases

100% traceability of fixed test decisions to source documents and spans
Leakage from records created after the event
Sensitivity to missing records and shortcuts based on age or diagnosis labels

Common synthetic checks

Evidence spans available at the time of the decision · 5/5 matched

Expected and observed values for these cases

What the official call asks

Covers community programme development and field validation in urban, rural and alcohol-use-disorder areas. The urban area specifies at least 400 participants and four or more institutions.

Proposal requirements, printed pages 53–56 · All twelve requirements

Interpretation and next evaluation

Recall in actual records requires a predefined population, observation window, timestamped independent reference and an eligible positive denominator.

Next check: Use structured event markers for synthetic evaluation. Define the requirements for longitudinal labels in external evaluation.

← All twelve projects