Objective: To enhance clinical decision-making in PD management by developing and validating a conformal prediction framework that forecasts Levodopa Equivalent Daily Dose (LEDD) changes with reliable uncertainty estimates, enabling data-informed medication adjustments

Background: Parkinson’s Disease (PD) medication management is complex due to variable disease progression and treatment response. Current approaches often lack systematic methods to predict needed dose changes and effectively use electronic health records (EHR) data. Levodopa, the gold-standard medication, requires careful dosing as inadequate or abrupt changes can cause complications like levodopa-induced dyskinesia. Our research applies conformal prediction, a statistical tool for uncertainty quantification, to forecasted LEDD changes, providing reliable prediction ranges that are guaranteed to contain the true medication needs for a pre-specified percentage of patients

Method: Using EHR data from UF Health Shands inpatient admissions (2011-2021), we leverage a two-step approach: first identifying patients likely to need medication changes, then predicting required LEDD adjustments. We evaluate three approaches: Naïve, Cross Validation (CV+), and Jackknife After Bootstrap (J+aB) to determine the most reliable method for clinical applications

Results: Our approach provides reliable predictions across different time periods. For short-term planning (6 months), we accurately predicted medication needs for 87% of patients, with dose changes within ±3.5% of current doses. Similar accuracy was maintained for one-year predictions (80% of patients, ±4% changes). For longer-term planning, predictions remained reliable but with wider ranges: two-year predictions (90% accuracy, ±9.5% changes) and four-year predictions (85% accuracy, ±13% changes), reflecting increasing uncertainty in medication needs over time

Conclusion: Through conformal prediction, we provide clinically valuable predictions for PD medication management with high certainty across different time horizons. Results align with clinical findings: providing precise estimates (±3.5-4%) for critical short-term adjustments and wider ranges (±9.5-13%) for long-term planning where uncertainty is higher. By leveraging EHR data with conformal prediction, we aim to provide evidence-based tools, bridging the gap between clinical and data-driven decision making to advance personalized medication management

Coverage & Interval Length comparison

Coverage & Interval Length through time

« Back to 2025 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/improved-decision-making-for-in-hospital-medication-management-in-parkinsons-disease/