Detection of Parkinson's Disease Through Static and Dynamic Spiral Test Drawings: A Transfer Learning Approach

Abstract

Parkinson's Disease detection can be considered a relevant yet overlooked issue in the field of research and medicine. Its effects are progressive in nature and worsens if not detected and treated accordingly. In this study, standardized tests such as static and dynamic spiral tests (SST and DST) are employed. On top of these, machine learning, specifically transfer learning is implemented. 14 pre-trained models are considered; 3 solvers are evaluated for each machine - these processes are repeated in 4 different scenarios. Based from the results, the pre-trained model with the highest accuracy is MobileNetV2 (93.94%), while the model with the sub-optimal performance is Vgg-19 (27.27%). In addition, it is realized that Stochastic Gradient Descent with Momentum (sgdm) and Adaptive Momentum (adam) are preferred over Root Mean Square Propagation (rmsprop) as the main solver for this kind of PD classification. Nonetheless, it is claimed that DST images are more correlated and significant than SST or a combination of both.