Early Detection & Stage Classification of Parkinson’s Disease using Deep Learning /
Muhammad Muzzamil Zeeshan
- 38p. ; 30cm.
Parkinson’s disease (PD) is caused by a lack of dopamine production by the substantia nigra in the brain. It is an enduring disorder without any cure, making it a burden on the patient and the society. PD is a complex disorder marked by many physical and non-physical manifestations, which differ for everyone. Clinicians might misdiagnose, waste time and resources to get a patient’s diagnosis or do not have enough expertise to diagnose a patient. Deep learning models tend to overfit with new data; thus, to prevent variance in the model, merging outputs has been proven effective. This study proposes an ensemble deep learning model, to automate PD detection and stage classification, which can handle different data by combining rules. The ensemble model (TransConvNet) links two state-of-the art deep learning models in decision level ensembling. The outputs are fused using averaging voting. Both neural networks utilize gait data provided by Physionet. The validation accuracy for PD detection reached 82%, while for PD stage classification, it reached 73%. This model delivers competitive and top-notch performance for severity and detection prediction for PD using gait. This can be used as a tool for PD detection or monitoring its development. Future work might include addition of models for better performance and reducing the training time of the models.