A hospital wants to predict, at the moment of discharge, whether a patient will be readmitted within 30 days. The candidate features are:
- Age and primary diagnosis
- Length of stay
- Number of prior admissions in the past year
- Number of follow-up clinic visits in the 30 days after discharge
Which feature must be removed before training, and why?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
The prediction is made at discharge, so anything that happens afterwards is unavailable when the model runs — and post-discharge visits are often caused by the readmission itself. That is target leakage: the model looks brilliant in validation and useless in production. The rule: every feature must be knowable at the moment of prediction.
Age raises a fairness question worth auditing, but it is known at discharge and does not leak the outcome. Fairness and leakage are different problems.
Features are allowed to correlate with each other. Length of stay is known at discharge and is legitimately predictive.
Prior admissions is exactly the kind of known-in-advance signal a readmission model should use. It is history, not the future.