A Machine Learning Framework for Parkinson's Disease Detection Through Turning Metrics.

This study presents a machine learning framework for the automated detection of Parkinson's disease (PD) using turning metrics derived from inertial measurement units (IMUs) and pressure sensors. Utilizing the Smart-Insole Dataset, which includes data from 29 individuals (13 healthy adults, 9 older adults, and 8 PD patients both on and off medication), the study introduces an algorithm to automatically detect turning points, eliminating the need for manual annotation or clinician supervision. A feature extraction pipeline is then implemented to extract entropy and smoothness metrics from the identified turns. These features are used to train Support Vector Machines (SVM), Random Forests (RF), and Gradient Boosting (GB) classifiers. The SVM model, combined with Chi-Squared feature selection, achieved the highest performance, reaching an F1-score of 92.92% for a 5-second time window. The RF model achieved 84.58% accuracy in the 3-second window, while the GB model attained 87.08% accuracy in the 4-second window. Feature selection analysis revealed that 64% of the top features were derived from angular velocity (40%) and acceleration (24%), with Bubble Entropy emerging as the most informative metric. The novelty of this framework lies in its fully automated, objective, and unsupervised diagnostic approach. The results demonstrate strong potential for enhancing early PD detection, supporting more accurate clinical diagnoses, improving decision-making processes, and contributing to personalized patient management in real-world environments.