Title: CASCADE Conformal Prediction: Uncertainty-Adaptive Prediction Intervals for Two-Stage Clinical Decision Support

Abstract: Managing medication for Parkinson’s Disease (PD) is a complex undertaking, complicated by the disease’s heterogeneous progression, inconsistent patient responses, and potential side effects. Although AI models are capable of forecasting the levodopa equivalent daily dose (LEDD)—a key metric for medication requirements—traditional uncertainty quantification methods often fall short. These standard approaches fail to distinguish the reliability of predictions, effectively treating clinical decisions with high confidence the same as those with low confidence.

To address this, we present CASCADE (Calibrated Adaptive Scaling via Conformal And Distributional Estimation), an innovative conformal prediction framework designed to propagate epistemic uncertainty from a screening classifier to refine downstream predictions. In contrast to conventional conformal techniques that depend on auxiliary residual regression, our method utilizes epistemic uncertainty derived from a primary classification task—specifically, determining whether a medication adjustment is necessary—to dynamically adjust the prediction intervals of a secondary regression task, which estimates the magnitude of that adjustment.

By translating Venn-Abers multi-probabilistic uncertainty directly into non-conformity scores, our framework enables continuous risk adaptation. We show that this "cascade effect" generates highly efficient intervals for patients with high confidence, resulting in intervals that are 38.9% narrower than those produced by standard conformal baselines. Simultaneously, the system automatically expands these intervals for uncertain cases to guarantee robust coverage. This approach bridges the divide between discrete clinical decision-making and continuous dose forecasting in PD management.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC