Diagnosis of Parkinson’s Disease Using Multimodal Deep Convolutional Network Based on Gait Analysis

Abstract

The neurodegenerative disease identified as Parkinson’s disease (PD) seriously impacts motor abilities and gait. Traditional gait evaluation methods frequently depend on subjective interpretation and clinical knowledge, that delays diagnosis. In this work, we utilize multimodal time-series data from smart insoles to offer a deep learning-based framework for objective and early PD detection. Gait sequences are sliced into 15-second intervals at a sampling rate of 100 Hz using the Smart Insole Dataset v1.0, which contains Inertial Measurement Unit (IMU) data—pressure, acceleration, and angular velocity signals from both feet. To capture temporal dependencies and transition dynamics, these segments are converted into Gramian Angular Field (GAF) and Markov Transition Field (MTF) images. For each modality, a multimodal convolutional neural network (MMCNN) is developed using parallel Conv2D layers, followed by max pooling and batch normalization. To enhance discriminative pattern learning, a Channel Attention Module is used to fuse and refine the extracted features. The model achieves high classification performance by classifying subjects into three groups: PD patients, elderly people (EL), and healthy adults (S).