A Study to Find Affordable AI Techniques for Early Parkinson's Disease Detection

Abstract

Parkinson's disease (PD) is a neurological degenerative disease that causes various movement impairments. In the context of statistical computer science, various supervised machine learning techniques are available in the previous studies. Previous studies focused on tabular features curated from the sensor data, speech, and images and emphasized manual feature engineering. This study focused on applying CPU-based deep learning techniques on drawings and speech signals and has analyzed the PD drawing classification results with various explainable AI and unsupervised learning algorithms. This study achieved state-of-the-art performance in drawing classification with GhostNet at a false alarm rate=0%, AUC=0.964 and achieved 3870 images per second inference speed. The novel convolutional half-auto-encoder is proposed for representation learning. For real-time inference, the LSTM model has archived 40 speech sample per minute inference rate.