Business Context

NorthStar Health operates 18 hospitals and wants to identify high-risk inpatients early enough to trigger care management interventions. You need to build a model that predicts whether a patient will experience an adverse outcome within 30 days of admission.

Dataset

You are given a de-identified hospital dataset built from electronic health records for adult inpatient admissions.

• Size: 142K admissions across 24 months, 54 features
• Target: Binary label indicating adverse outcome within 30 days of admission (ICU transfer, in-hospital mortality, or readmission)
• Class balance: 11.6% positive, 88.4% negative
• Missing data: 18% missing in some lab features, 6% missing in vitals, and sparse missingness in socioeconomic fields

Success Criteria

A good solution should achieve strong ranking performance and clinically useful recall. Target ROC-AUC >= 0.84, PR-AUC >= 0.42, and recall >= 0.75 at a threshold that keeps precision practical for care teams.

Constraints

• Predictions must be generated within 200 ms per patient in batch scoring
• Clinicians need interpretable drivers of risk
• Retraining should be feasible monthly using standard Python ML tooling
• Avoid leakage from post-outcome information

Deliverables

1. Select and justify candidate algorithms for this patient outcome prediction task
2. Build a training pipeline with preprocessing, feature engineering, and validation
3. Compare at least one interpretable baseline and one stronger nonlinear model
4. Choose evaluation metrics appropriate for imbalanced clinical classification
5. Describe how you would deploy, monitor, and retrain the model in production