Assisting Physician Diagnosis of Parkinson’s: Predicting the Disease by Avoiding Subjective Tests

This case study aims to predict Parkinson’s disease using curated data from the Micheal J. Fox Foundation. The goal is to assist physicians by creating a reliable and reproducible tool through machine learning models. Unique to this paper, commonly used subjective predictors like the UDPRS are removed in favor of non-biased objective metrics. Data consisted of demographic, clinical, biological, and genetic subgroups. Validation through training and test splits reveal Random Forest as the best classifier of the disease. A model accuracy of 85% beats current estimations of physician diagnosis. Model predictors are relatively accessible and support the function of a rudimentary initial disease screening. Caudate region-related variables from DaTscan data heavily influence model prediction while BMP derivatives offer further research opportunities. Steps forward involve medical integration and advisement from diagnosing physicians. Alternative designs consider time-series modeling.