Advancing Parkinson’s disease detection through multi-dimensional machine learning: a comprehensive framework using wearable movement sensor analytics

TL;DR

This study explores how wearable sensors and machine learning can improve the detection of Parkinson’s disease by analyzing movement data and identifying key features that distinguish PD patients from healthy individuals.

Contribution

The study introduces a comprehensive framework combining multiple machine learning algorithms, feature types, and optimization techniques for accurate and interpretable PD detection using wearable sensor data.

Findings

Random Forest with Particle Swarm Optimization achieved the highest PD detection accuracy (87.65%).

Statistical features were most influential, while entropy-based measures and standard deviations were key predictors.

Accelerometer-derived complexity features strongly indicated PD, while gyroscope features were more relevant for non-PD cases.

Abstract

wearable movement sensor technology shows promise for objective assessment of Parkinson’s disease (PD) motor symptoms, but optimal machine learning approaches and feature sets for accurate PD detection remain unclear. This study provides a comprehensive evaluation of classification algorithms, feature contributions, and optimization techniques for PD detection using wearable movement sensor data.

We compared twelve diverse machine learning classifiers on motion sensor data, conducted systematic feature ablation studies across statistical, frequency-domain, dynamic, and complexity feature categories, optimized Random Forest parameters using three meta-heuristic algorithms, which is Particle Swarm Optimization (PSO), Improved Satin Swarm Algorithm (ISSA), and Enhanced Whale Optimization Algorithm (EWOA), and performed SHAP value analysis to identify the most influential features and their impact patterns.

Random Forest demonstrated superior performance (86.7% accuracy) among all classifiers. Statistical features contributed most significantly to classification performance, while complexity, dynamic, and frequency domain features provided complementary information. PSO-optimized Random Forest achieved 87.65% accuracy, outperforming other optimization approaches. SHAP analysis identified entropy-based measures and standard deviations as the most influential features, with accelerometer-derived complexity measures driving high-probability PD predictions and gyroscope-derived measurements dominating low-probability outcomes.

Ensemble-based methods effectively capture the complex, non-linear relationship between movement characteristics and PD diagnosis. Comprehensive feature extraction frameworks incorporating multiple movement dimensions significantly enhance detection accuracy. The asymmetric feature influence patterns for positive versus negative predictions align with clinical understanding of PD as a disorder characterized by altered movement complexity and variability. These findings provide a foundation for developing accurate, interpretable wearable monitoring systems for Parkinson’s disease detection and management.

Figures