Classification of hyperkinesia in Parkinson patients using mobile sensors

Abstract: In this thesis, we explore the possibility to monitoring hyperkinesia in people who suffer from Parkinson's disease (PD) using sensors in mobile phones. This is done by first collecting data from 25 patients diagnosed with PD by using a sensor-recording smartphone in a bag attached to the stomach, and at the same time let trained professionals make assessments of the degree of which they show signs of hyperkinesia, on the clinical dyskinesia rating scale, CDRS. Given the labels and the sensor data, a set of models has been trained. Both models for binary classification, i.e., predicting the presence of hyperkinesia vs no hyperkinesia, as well as models aiming to estimate the CDRS score were investigated. As the available data is a mere 429 samples, a key part of the work has been to self-engineer features descriptive to signs of hyperkinesia. The proposed models are kernel support vector machines for both the binary classification and for the regression. The proposed method provides results that are in line with what can be expected of an assessment by a trained professional.