ic_001 · Industry: Healthcare — 90-day readmission prediction

Status: ✓ Passing · Last run: 2026-06-07T06:00:33 · Pipeline: v2 · History: 1/2 runs passed (50%)

What was tested

Canonical healthcare ML problem: predict whether each patient will be hospitalized again within 90 days, using patient demographics + diagnoses + encounters + medications + lab results, with a quarterly observation panel for the target. Mirrors the actual customer shape Yifan was iterating on for 6 days. Six tables (5 entities/events + 1 labels panel). Planted signal: high-comorbidity patients with frequent ER encounters + abnormal labs have higher readmission probability.

What we planted in the data

• 300 entities
• Signal strength: 0.70 (sigmoid slope multiplier)

What the system did

• Training rows built: 1,200
• Features used in model: 25
• Model selected: RandomForestClassifier
• CV AUC: 0.745

Top features by importance:

• diagnoses_count_365d (0.150)
• diagnoses_count_180d (0.141)
• bnp_avg_90d (0.123)
• all_encounters_count_180d (0.113)
• inpatient_visits_count_365d (0.109)

What we checked — all assertions passed

How this could have gone wrong

• If no_pipeline_errors failed: Pipeline must run cleanly end-to-end through 6 tables.
• If feature_survival_rate_above failed: At least 40% of generated features must survive dry-run.
• If source_tables_contributing_above failed: At least 3 of the 5 source tables must contribute surviving features.
• If auc_in_band failed: AUC at least meaningfully above random.

Why this case matters

This shape (entity + multiple event streams + panel labels) is THE canonical healthcare ML problem. If we can do this end-to-end the buyer-facing pitch is concrete: "we just trained a 90-day readmission model from 6 tables in 10 minutes."

Reproducing

# from auto_insight_api/
python -m validation.v2 run ic_001 --pipeline v2 -v

• Case config: validation/v2/cases/ic_001_healthcare_readmission.yaml
• Data shape: industry_healthcare_readmission
• Analytics type: ml_binary

← back to validation report