Research Question:
How well do predictive models for health outcomes perform in terms of target sharpness?
Objectives:
Introduce Target Sharpness as a complementary evaluation metric and apply Negative Control Outcomes to penalize overly broad predictiveness, ensuring models are finely tuned to the intended outcome.
Scientific Rationale:
Traditional health prediction models prioritize metrics like AUC, accuracy, and calibration. Yet, a model can perform well on these metrics while still lacking target sharpness-the ability to distinguish the specific outcome of interest from unrelated but correlated events. This project aims to improve model specificity by identifying truly relevant features and promoting precise, interpretable predictions for clinical and epidemiological use.
