Feature Extraction and Classification of Medication-Induced Hyperkinesia During Treatment of Parkinson's Disease

Abstract: The recent progress in wearable sensor technology, signal processing and machine learning enables novel applications in fields such as automatic disease symptom tracking and classification. In this project, data from an ongoing Parkinson’s disease study at the Lund University Hospital is gathered and analyzed, and classification models of varying complexity for symptom severity are evaluated. The accelerom- eter and gyroscope data is provided by a single mobile phone fastened around each patient’s trunk and is complemented by a certified specialist’s score labels, in clin- ical dyskinesia rating scale (CDRS). Out of four available tasks performed by the patient during measurement, the two tasks “walking” and “sitting and describing a picture” were cut out and analyzed in-depth. The most practically usable simulation scenario was found to be forming in- dividual models for the patients, with the best performing model proving to be an unsupervised feature extracting autoencoder combined with a linear discriminant. We could with descent accuracy distinguish between whole test signals in a binary- class scenario for a majority of patients and perform skillfully in the multi-class scenario, although not well enough for practical usability. The total patient-average, mean macro f1-score and accuracy obtained for binary-class classification of short (2 second) signal segments from the “describe picture”-task were 0.74 and 0.81 re- spectively. The corresponding mean macro f1-score and accuracy for the multi-class classification case were 0.52 and 0.65 respectively.