A Simple and Reproducible Machine Learning Pipeline for Parkinson’s Disease Detection Using Smartwatch-Based Inertial Signals

Abstract

Parkinson’s disease (PD) is a progressive neurological disorder characterized by motor symptoms such as tremor, rigidity, and bradykinesia. Wearable inertial sensors enable non-invasive and cost-effective assessment of motor abnormalities in real-world settings. Despite recent advances in deep learning, many existing approaches rely on complex architectures with limited interpretability and inconsistent evaluation protocols. This study proposes a simple and reproducible classical machine learning pipeline for PD detection using smartwatch-based inertial signals from the PADS dataset. Spectral and statistical features were extracted from accelerometer and gyroscope signals, and LASSObased feature selection was applied within a nested subject-level cross-validation framework to prevent data leakage. Several classifiers, including Logistic Regression, Support Vector Machine, Random Forest, CatBoost, and Multi-Layer Perceptron, were evaluated. The proposed pipeline achieved 79.26% balanced accuracy, 87.32% accuracy, and an F1-score of 0.92 for PD vs. healthy control classification, while 67.15% balanced accuracy was obtained for the more challenging PD vs. differential diagnosis task. Feature analysis showed that PD vs. healthy control discrimination is dominated by tremor-related spectral and amplitude features, whereas variability-related features are more relevant for differential diagnosis. These results demonstrate that competitive performance can be achieved using a simple and interpretable pipeline, providing a practical alternative to more complex deep learning approaches.