Business Context

MediScan Health operates urgent-care clinics that process roughly 12,000 patient visits per week. The clinical operations team wants a machine learning model that flags whether a patient is likely to have a target condition (for example, pneumonia) using structured intake, vitals, lab results, and symptom data to support physician triage—not replace clinical judgment.

Dataset

You are given a historical supervised learning dataset built from prior visits with confirmed diagnoses.

• Rows: 148,000 patient visits collected over 24 months
• Target: Binary label indicating confirmed diagnosis of the target condition within 48 hours
• Class balance: 14% positive, 86% negative
• Missing data: ~18% missing in lab features because not all tests are ordered; <3% missing in vitals and symptoms
• Data quality note: Multiple visits from the same patient may exist, so leakage must be avoided

Success Criteria

A good solution should achieve strong recall for true positive cases while keeping false alarms manageable for clinicians. Target performance is recall >= 0.85, precision >= 0.45, and ROC-AUC >= 0.88 on a held-out test set.

Constraints

• Predictions must be returned in <100 ms per patient in the triage application
• The model must be explainable enough for clinical review
• Training should use only structured tabular data
• The solution must minimize patient-level leakage across splits

Deliverables

1. Build a binary classification pipeline for automated diagnostic triage.
2. Explain model choice, preprocessing, and leakage prevention.
3. Evaluate the model with clinically relevant metrics and threshold tuning.
4. Provide feature importance or explanation outputs suitable for review.
5. Describe how you would deploy and monitor the model in production.