Class-Conditional Conformal Prediction with Uncertainty-Triggered Selective Explainability for HighStakes Imbalanced Tabular Decisions

Machine learning models deployed in high-stakes clinical and financial decision-making conventionally output uncalibrated point predictions and soft probabilities that exhibit unwarranted overconfidence. When applied to severely imbalanced tabular distributions—such as acute stroke events or rare cardiovascular disorders—traditional classification objectives incentivize models to minimize global risk by predicting the dominant majority class, inducing near-total sensitivity collapse on critical minority instances. While Conformal Prediction (CP) provides distribution-free finite-sample guarantees, standard marginal calibration suffers from an overlooked failure mode: it achieves nominal marginal coverage globally by over-covering the majority class while catastrophically under-covering the rare disease class (collapsing to coverage on stroke events). To resolve this safety-critical vulnerability, we introduce the Adaptive Penalized Class-Conditional Conformal Predictor (AP-CCCP), a mathematically rigorous framework establishing finite-sample class-conditional guarantees conditioned on exchangeability. AP-CCCP formulates a regularized non-conformity metric integrating an adaptive rarity penalty with margin epistemic uncertainty. Furthermore, we design an Uncertainty-Triggered Selective Explainability protocol that utilizes conformal set cardinality as an autonomous triage trigger: unambiguous singleton predictions () receive an automated fast-pass, whereas borderline or ambiguous cases () trigger on-demand TreeSHAP attribution maps. We benchmark across three clinical cohorts (Kaggle Stroke with 1:20 imbalance, Framingham Heart Study with 1:5.6 imbalance, and Pima Diabetes with 1:1.9 imbalance) and four heterogeneous model families (LightGBM, XGBoost, Random Forest, Deep MLP) across 120 stratified cross-validation folds. While Standard Split CP omits of true stroke cases, AP-CCCP rigorously secures rare-class coverage while reducing post-hoc explainability computational overhead by with automated triage precision up to . This establishes a robust foundation for trustworthy and computationally efficient tabular decision systems.