Efficient quantification of Parkinson's disease severity using augmented time-series data.

Abstract

The diagnosis of Parkinson's disease relies heavily on the subjective assessment of physicians, which depends on their individual experience and training, potentially leading to inconsistent diagnostic results. Therefore, developing an objective and efficient diagnostic method is essential to improve the accuracy and timeliness of Parkinson's disease diagnosis. In this study, we utilized the PhysioNet dataset, a time-series dataset comprising data from 93 Parkinson's patients and 73 healthy individuals. The dataset contains vertical ground reaction forces recorded from 16 sensors (8 per foot) during a 2-minute test at a sampling rate of 100 Hz. To address challenges such as limited dataset size, high labeling noise, and high intra-class variability, we performed data preprocessing and applied various data augmentation techniques, including jittering, scaling, rotation, permutation, magnitude warping, time warping, cropping, and linear residuals. These methods were evaluated using one-dimensional-convolutional neural network (1D-ConvNet) and one-dimensional Transformer networks. By conducting 10-fold cross-validation, we observed significant improvements in classification performance. The best data augmentation strategy achieved 90.8% accuracy, 92.0% precision, 91.0% recall, and a 91.0% F1 score in assessing disease severity. These results highlight the importance of selecting appropriate data augmentation techniques for time-series data to improve model generalization and diagnostic reliability, while also offering new insights for researchers working with sensor device data. Our results demonstrate that data-enhanced methods can significantly boost the performance of machine-learning models in the field of Parkinson's disease diagnosis.