Enhanced Parkinson's Disease Detection Through Multimodal Feature Fusion of Speech and Gait Signals Using an Optimized Ensemble Learning Framework

Abstract

Parkinson's disease (PD) is a progressive disease affecting motor and non-motor skills, making primary and correct diagnosis critical in its clinical organization. The current diagnostic techniques have shown limited success, leading to the development of a proposed multimodal machine learning-based method to correct the issue of diagnosing PD. In order to train our proposed model, two openly available benchmark datasets were used, namely, the UCI Parkinson’s Voice Dataset (195 audio recordings obtained from 31 PD patients and 23 controls) and PhysioNet Gait in Parkinson’s Disease Dataset (wearable inertial sensor data). A thorough preprocessing technique was used separately on both modalities. Features obtained from both modalities include voice features like jitter, shimmer, harmonics-to-noise ratio (HNR), and MFCC, and gait features such as stride, cadence, accelerations, and symmetries. PCA was performed for both modalities while considering 95% of the variance. The feature-level fusion process was next adopted for the purpose of combining the low-dimensional representations into one coherent feature vector. In this regard, the classification process was performed using an ensemble learning architecture that uses the three mentioned classifiers (RF, GB, and XGB) in conjunction with a majority vote scheme. From the investigational outcomes, it is obvious that a classification accuracy rate of 93% has been achieved along with other metrics, including precision (94.2%), recall (92.1%) and F1 score (93.1%). These outcomes show that the suggested solution is able to outperform any single-modality as well as individual classifier solutions.