Machine Learning Models for the Prediction of Levodopa Response to Tremor in Parkinson's Disease

Fangfei Li^1,2Shinuan Lin^3,4Rui Yan^5,6Yusha Cui^5,6Kang Ren^3,4Zhonglue Chen^3,4*Lingyan Ma^5,6*Tao Feng^6,7*

• ¹Center for Movement Disorders, Department of Neurology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China, beijing, China
• ²Beijing Chaoyang Hospital Affiliated to Capital Medical University, Beijing, China
• ³GYENNO SCIENCE CO., LTD., Shenzhen, China, shenzhen, China
• ⁴HUST-GYENNO CNS Intelligent Digital Medicine Technology Center, School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, wuhan, China
• ⁵Beijing Tiantan Hospital Department of Neurology, Beijing, China
• ⁶China National Clinical Research Center for Neurological Diseases, Beijing, China
• ⁷Department of Neurology, Beijing Chaoyang Hospital, Capital Medical University, Beijing, China, beijing, China

Objectives To develop and validate machine learning models to predict levodopa responsiveness of tremor in Parkinson's disease (PD) patients. Methods: A total of 197 PD tremor patients underwent Levodopa Challenge Tests and were classified as having levodopa-responsive or levodopa-resistant tremor. Clinical and electromyogram (EMG) tremor analysis variables were recorded. The dataset was randomly divided into a training set (80%) and a test set (20%). To distinguish between the two groups, Support vector machine (SVM), random forest (RF), and logistic regression (LR) models were developed using training data. The optimal model was validated on test data. Calibration and decision curve analyses assessed model reliability and clinical utility. Results Among 197 patients, 95 had levodopa-responsive tremor and 102 had levodopa-resistant tremor. The SVM model showed the best performance, achieving an accuracy of 81.5% in five-fold cross-validation, with a Kappa score of 0.624, sensitivity of 84.3%, specificity of 77.9%, and an area under the curve (AUC) of 0.850. Performance remained consistent on test data, with 82.5% accuracy, 0.653 Kappa, 93.8% sensitivity, and 75% specificity and 0.896 AUC. The best model incorporated 6 predictors: resting tremor score, rigidity/tremor ratio, postural and kinetic tremor score, disease duration, the Movement Disorder Society's Unified Parkinson's Disease Rating Scale III (MDS-UPDRS III) /disease duration, supine diastolic blood pressure (DBP). Conclusion The SVM model, incorporating 6 key indicators, holds significant potential for predicting levodopa responsiveness in PD tremor, offering a valuable tool for the precise treatment of tremor in PD patients.