Monitoring and Quantification of Parkinson’s Disease Motor Symptoms using Wearable Sensors for Smart Healthcare

Abstract

Research on Parkinson’s disease has been ongoing for a while. Since there is now no known treatment for it, there are occasionally some symptoms that can be relieved with particular medications, surgical procedures, or other therapies. Tremors, dyskinesia, and bradykinesia are the three most noticeable signs of Parkinson’s disease. The detection and prediction of these symptoms is extremely important since it can aid in preventing accidents, injuries, etc. by producing notifications. Our study focuses on the inertial data-based detection and categorization of these Parkinson’s anomalies. In the levodopa trial, inertial data was gathered from three separate body regions, including the upper two limbs and the waist, using three different sensors named Geneactiv and Pebble Smartwatch and Samsung Galaxy S2 smartphone. We present a two-step pipeline for anomaly detection and classification, with the first step (binary classification) classifying an input signal sequence into two categories: abnormal or normal. In the second step (multi-label classification), if the input signal is abnormal, we classify the different types of abnormalities (tremors, dyskinesia, and bradykinesia) that have been detected in the signal being analysed. In this two-step process for anomaly detection and classification, we proposed a Transformer model with the following settings : During training, the grid approach was utilised. The following hyperparameters were set for the model: epochs were set to 150, number of layers to 3, embedded layer size to 128, fully connected layer size to 256, number of attention heads to 6, dropout to 0.1, attention dropout to 0.1, optimizer was Adam’,learning rate’ to 1e-3, warm-up steps to 10 and batch size to 128. Furthermore, the algorithm used preprocessing techniques like data augmentation (Time shift augmentation, Scaling augmentation, Rotation augmentation, and Noise injection augmentation), pairwise sampling, and downsampling to improve the model’s performance and ability to detect anomalies accurately. We have tested the proposed algorithm using publicly available datasets and achieved an accuracy of 97.86% for the Binary Class with a batch processing inference time of 0.04 seconds and GPU utilisation of 62.24%, GPU memory allocation of 10.04%, and disc utilisation of 30.2%. The accuracy for multiclass is 98.75%, with a batch processing inference time of 0.06 seconds and GPU utilisation of 71.3%, GPU memory allocation of 12.48%, and disc utilisation of 31.8%. When compared to state-of-the-art methodologies, these statistics show that our developed algorithm is light weight, and the high accuracy underlines the system’s ability in correctly classifying distinct types of anomalous occurrences linked with Parkinson’s disease.