AI-Based Risk Prediction Models

Abstract

Risk prediction models -- systems that estimate the probability of adverse future events from observed covariates -- are among the most consequential applications of AI, influencing decisions in clinical care, financial lending, insurance underwriting, criminal justice, and infrastructure management. The stakes of getting risk prediction right are asymmetric: underestimating risk can lead to catastrophic outcomes (missed disease diagnoses, undetected loan defaults, infrastructure failures), while overestimating risk causes unnecessary interventions, denied opportunities, and potential discrimination against individuals and groups. This study evaluates ten AI risk prediction models across five high-stakes domains: 30-day hospital readmission, credit default, surgical complication risk, wildfire spread probability, and bridge structural failure. Models evaluated include logistic regression, gradient boosting (XGBoost, LightGBM), neural networks (TabNet, FT-Transformer), survival models (DeepSurv), fairness-constrained learning (FairGB), uncertainty-aware models (Conformal Risk, Bayesian NN), and a proposed Calibrated Risk Ensemble (CRE). Evaluation covers discriminative performance (AUC, AP), calibration (ECE, Brier score), fairness (demographic parity, equalised odds), and decision-theoretic utility (net benefit at clinically relevant thresholds). CRE achieves the best mean AUC (0.864) and best calibration (ECE 0.028) across all domains. FairGB achieves the best fairness-accuracy balance. Conformal Risk prediction achieves guaranteed coverage at user-specified risk thresholds. A risk model quality framework covering all four evaluation dimensions is proposed.