Predicting Wearing-Off in Parkinson’s Disease Patients Using Multimodal Time-Series Data and Ensemble Learning

Parkinson’s disease (PD) is a neurodegenerative disorder that mainly affects the movement of patients with PD. PD affects not only motor functions but also non-motor functions, highlighting its extensive and multifaceted impact. Patients with PD take designed treatment mainly with anti-PD drugs. When the effects of anti-PD drugs diminish, patients start to experience PD symptoms that persist until the next dose, and this phenomenon refers to “wearing off” (WO). WO refers to the phenomenon when anti-PD medicine lose their effect. Over time, WO episodes become shorter and more frequent. This study aims to predict the wearing-off of anti-PD medication using multimodal timeseries data and symptom analytics. To address class imbalance and enhance prediction performance, we incorporated data augmentation techniques, specifically Synthetic Minority Oversampling Technique (SMOTE) and Adaptive Synthetic Sampling (ADASYN), to balance the dataset. Our predictive framework employs an ensemble learning approach that integrates a Neural Network (NN), Random Forest (RF), and XGBoost (XGB) models, using fitness tracker data and self-reported symptoms from a smartphone app in a real-world setting. Through robust data preprocessing, augmentation, and ensemble modeling, our approach achieved an average accuracy of $98.06 \%$, precision of $\mathbf{8 5. 8 3 \%}$, recall of $\mathbf{9 0. 5 5 \%}$, and an F1-score of $\mathbf{8 7. 6 5 \%}$ - significantly outperforming the state-of-the-art methods, including leading solutions from the global PD WO competition (accuracy of 93.2%, at Intl. Conf. on Activity and Behavior Computing, 2023). The predictive model has the potential to help clinicians develop more personalized and precise treatment strategies for people affected by PD. Effective management of symptoms and episodes of WO can improve the ability of people with PD to perform daily activities with greater ease and comfort.