Kinect Enabled Deep Learning Model for Parkinson’s Gait Analysis

Detection of Parkinson s disease (PD) at an early stage and reliably is a serious challenge as it is a progressive, and subtle condition with gait impairments. Due to lack of reliable methods that effectively utilize the available costly equipments and are invasive techniques or requiring painstaking techniques of feature extraction. To bridging such a gap, this paper present a Kinect based hybrid deep learning model that combines 1dimensional convolutional neural networks (Conv1D) with long short-term memory (LSTM) units to automatically analyze gaits. The data collected in walking sequences is non-invasive three-dimensional skeletal joint data using the Microsoft Kinect sensor because of the unique feature of Kinect sensors that allows a precise tracking of gait dynamics. Conv1D layers are used to learn meaningful spatial-temporal characteristics of raw sequences, and LSTM units to represent long-term structure and progressive changes in movement patterns. The training is performed with a supervised learning strategy based on categorical cross-entropy loss and Adam optimization. Performance is measured in terms of accuracy, plots of loss over time and confusion matrix. The experimental findings illustrate an accuracy rate of $89-92$ percent, a high level of sensitivity in the identification of the PD cases highlighting the potential of the system as a low-cost and real-time application in early diagnosis, constant monitoring, and integration into both the clinical and home-based health facilities.