Mental-health inpatient readmission prediction

Decision and workflow

This system scores fictional mental-health inpatient discharge events for a 28-day return and supports transition planning. It is not a diagnosis, a clinical recommendation, or a substitute for follow-up planning by qualified care teams.

System approach

The workflow uses a patient-disjoint temporal split and compares prevalence/rules, logistic regression, and boosted trees. The review queue presents probability bands, perturbation-based reason codes, missing-data cautions, capacity controls, and an explicit human-review boundary.

Baseline and evaluation

Evaluation covers PR-AUC, Brier score, calibration, review-capacity recall, subgroup behavior, failure cases, and sensitivity to label definitions. Planned and unplanned returns, incomplete network follow-up, and post-discharge workload are treated as modelling decisions that must stay visible.

Operational boundary

The queue supports a reviewer deciding where transition-support attention may be useful. It does not label a person as high risk in a clinical sense or imply that an intervention will change an outcome. Small groups are suppressed and limitations are shown alongside the visuals.

Reproducibility

The project includes synthetic cohort generation, data cards, model metadata, demo records, evaluation and monitoring reports, failure analysis, and a simulated documentation-drift incident with response thresholds.

Selected evidence

Each visual answers a specific question about system behavior. Read the caption and limitation together.

Readmission capacity result showing recall and precision at the chosen review workload — **Hero result**Capacity is the operational decision: the queue shows what a fixed review workload captures under the synthetic cohort. *Limitation: The cohort and outcomes are simulated.*

Discharge index and inclusive twenty-eight-day readmission label window — **Cohort and label window**The index date and inclusive label window make the leakage boundary visible. *Limitation: The network and follow-up context are fictional.*

Recall and precision as transition-support review capacity changes — **Capacity analysis**Review capacity is treated as a design input rather than an afterthought. *Limitation: Capacity is an assumption for this portfolio study.*

Failure analysis showing false positives, false negatives, missing-data context, and threshold errors — **Failure analysis**Failure cases connect threshold behavior to missing follow-up context and reviewer interpretation. *Limitation: The errors are generated from synthetic records.*

Simulated monitoring dashboard showing documentation drift and response thresholds — **Monitoring**The monitoring view follows a documentation-drift incident through detection and response. *Limitation: The incident is injected and simulated.*

← All work Next: Customer-interaction RAG agent ↗