You already know: classification, calibration and leakage (D3), causal inference and uplift (D10) and, from the churn case, that predicting who is at risk is not the same as knowing whom you can help. This case adds clinical decision support: data from health records with messy timing, labels that are hard to define, strict privacy, clinicians who must stay in charge, and an intervention whose effect must be proven. The design questions: how to define the label and avoid leakage, whom to target with fixed capacity, how the score enters the care team's workflow, and how to show the programme works.
How to read this lesson. No code, on purpose. Numbers marked assumed are assumptions you state out loud. This is a systems design case, not clinical guidance.
1. The prompt, and the questions that change the design
A mid-level answer: "train a model on health-record data and flag patients above a threshold". A staff candidate first asks what the follow-up programme does and how many patients it can take, how readmission is defined, what data is known at the moment of scoring, and where the care team will see the score. Those answers produce capacity-aware targeting, an unplanned, any-hospital, 30-day label, point-in-time features, and scores inside the worklist the team already uses.
Assumptions you state: