Evolving classifiers to inform clinical assessment of Parkinson's disease

We describe the use of a genetic programming system to induce classifiers that can discriminate between Parkinson's disease patients and healthy age-matched controls. The best evolved classifer achieved an AUC of 0.92, which is comparable with clinical diagnosis rates. Compared to previous studies of this nature, we used a relatively large sample of 49 PD patients and 41 controls, allowing us to better capture the wide diversity seen within the Parkinson's population. Classifiers were induced from recordings of these subjects' movements as they carried out repetitive finger tapping, a standard clinical assessment for Parkinson's disease. For ease of interpretability, we used a relatively simple window-based classifier architecture which captures patterns that occur over a single tap cycle. Analysis of window matches suggested the importance of peak closing deceleration as a basis for classification. This was supported by a follow-up analysis of the data set, showing that closing deceleration is more discriminative than features typically used in clinical assessment of finger tapping.