Machine-learning model for the prediction of acute orthostatic hypotension after levodopa administration

Levodopa could induce orthostatic hypotension (OH) in Parkinson's disease (PD) patients. Accurate prediction of acute OH post levodopa (AOHPL) is important for rational drug use in PD patients. Here, we develop and validate a prediction model of AOHPL to facilitate physicians in identifying patients at higher probability of developing AOHPL.

The study involved 497 PD inpatients who underwent a levodopa challenge test (LCT) and the supine-to-standing test (STS) four times during LCT. Patients were divided into two groups based on whether OH occurred during levodopa effectiveness (AOHPL) or not (non-AOHPL). The dataset was randomly split into training (80%) and independent test data (20%). Several models were trained and compared for discrimination between AOHPL and non-AOHPL. Final model was evaluated on independent test data. Shapley additive explanations (SHAP) values were employed to reveal how variables explain specific predictions for given observations in the independent test data.

We included 180 PD patients without AOHPL and 194 PD patients with AOHPL to develop and validate predictive models. Random Forest was selected as our final model as its leave-one-out cross validation performance [AUC_ROC 0.776, accuracy 73.6%, sensitivity 71.6%, specificity 75.7%] outperformed other models. The most crucial features in this predictive model were the maximal SBP drop and DBP drop of STS before medication (ΔSBP/ΔDBP). We achieved a prediction accuracy of 72% on independent test data. ΔSBP, ΔDBP, and standing mean artery pressure were the top three variables that contributed most to the predictions across all individual observations in the independent test data.

The validated classifier could serve as a valuable tool for clinicians, offering the probability of a patient developing AOHPL at an early stage. This supports clinical decision-making, potentially enhancing the quality of life for PD patients.