Objective: To develop and validate a uncertainty-adaptive AI framework to improve the safety and clinical utility of levodopa dose predictions by explicitly flagging high-risk, ambiguous cases that warrant closer clinical review.

Background: Medication adjustments in Parkinson’s Disease (PD) require careful assessment: too little risks untreated motor symptoms, too much risks levodopa-induced dyskinesia or other complications. This balance is especially challenging in patients with motor fluctuations, where the cost of a miscalibrated recommendation impact patient’s lives. Current AI tools carry a hidden risk: expressing equal certainty for straightforward and clinically ambiguous cases alike. We develop CASCADE (Calibrated Adaptive Scaling via Conformal And Distributional Estimation) to close this gap, building a system that signals its own uncertainty and adjusts dosing recommendations accordingly.

Method: CASCADE was developed and tested on a cohort of 631 PD patients from UF Health, with over 10 years of longitudinal clinical data. The system first assesses whether a medication change is needed, then uses the patient’s epistemic uncertainty to scale the range of its dose recommendation, narrowing it when the evidence is clear, widening it when it is not. A tunable parameter (β) allows the sensitivity of this behavior to be adjusted for different clinical settings, ensured through conformal prediction guarantees.

Results: CASCADE achieved fourfold greater adaptivity versus standard AI baselines (Cascade Ratio: 4.27). For high-risk patients, CASCADE expanded prediction intervals by 150.7% and achieved 92.16% coverage, urging caution in dose adjustments where over-treatment risks levodopa-induced dyskinesia or other complications. For low-risk patients, medication ranges narrowed by 41.2% while maintaining 72.82% coverage, supporting confident clinical decision-making without sacrificing safety.

Conclusion: Current AI dosing tools are precise where they need to be cautious. By demonstrating that uncertainty-adaptive predictions meaningfully reduce the risk of false precision in high-stakes dosing decisions, this work sets a new standard for what clinically responsible AI should look like in movement disorders. Validated on a large real-world PD cohort, these findings make a strong case that uncertainty-adaptive design is a prerequisite, not an enhancement, for AI-assisted medication management in Parkinson’s disease.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/safer-ai-dosing-recommendations-in-parkinsons-disease-through-uncertainty-adaptive-prediction/