Systematic detection of clinically inverted feature effects

Develop an established method for systematically detecting clinically implausible reversals in which machine-learning models associate worse disease characteristics with lower predicted risk.

Background

The paper shows that unconstrained XGBoost, logistic regression, and random-forest models can learn associations opposite to established medical knowledge, such as assigning lower bladder-cancer recurrence risk to higher tumour stage or carcinoma in situ. These reversals were not detected by discrimination, calibration, or SHAP-based explanations. The authors present their counterfactual direction test as a framework for detecting such failures, but frame the broader methodological gap as the absence of an established systematic detection method.