Unsupervised Learning from Motion Sensor Data to Assess the Condition of Patients with Parkinson’s Disease

Abstract

Parkinson’s disease (PD) is a chronic neurodegenerative disorder that predominantly affects the patient’s motor system, resulting in muscle rigidity, bradykinesia, tremor, and postural instability. As the disease slowly progresses, the symptoms worsen, and regular monitoring is required to adjust the treatment accordingly. The objective evaluation of the patient’s condition is sometimes rather difficult and automated systems based on various sensors could be helpful to the physicians. The data in this paper come from a clinical study of 19 advanced PD patients with motor fluctuations. The measurements used come from the motion sensors the patients wore during the study. The paper presents an unsupervised learning approach applied on this data with the aim of checking whether sensor data alone can indicate the patient’s motor state. The rationale for the unsupervised approach is that there was significant inter-physician disagreement on the patient’s condition (target value for supervised machine learning). The input to clustering came from sensor data alone. The resulting clusters were matched against the physicians’ estimates showing relatively good agreement.