Home Monitoring of Drug Response in Patients with Parkinson's Disease using Wearable Sensors

Christoph Matthias Kanzler, Sunghoon Ivan Lee1, Jean-Francois Daneault, Fatemeh Noushin Golabchi, Julius Hannink2, Cristian Federico Pasluosta, Bjoern M Eskofier2, Paolo Bonato3

  • 1UMass Amherst
  • 2Friedrich-Alexander-Universität Erlangen-Nürnberg
  • 3Harvard Medical School

Details

14:15 - 14:30 | Thu 27 Oct | Main Auditorium | ThCT1.4

Session: Technical Session 5: Enhancing Gait and Movement in Neurological Conditions

Abstract

Parkinson's disease (PD) is the second most common neurodegenerative disorder. Tremor is one of the main motor symptoms associated with PD. Wearable sensors enable unobtrusive and mobile assessment of motor symptoms during activities of daily living, which has potential to improve the management of PD. The goal of this study was to investigate the use of wearable sensors to predict the level of tremor and its response to medication in patients' home setting. Fourteen PD patients were monitored continuously over four days using wearable sensors (accelerometers), placed bilaterally on the ankles and forearms. On the first and fourth day, patients performed specific motor tasks: sitting, finger to nose, and alternating hand movements in the laboratory setting. The motor tasks were repeated six times at 30 min intervals. The severity of tremor during these tasks was labeled by a physician based on the UPDRS scale from 0 ("no tremor") to 4 ("severe tremor"). On the second and third day, data were collected in patients' home setting. Patients performed their regular daily activities and followed their regular medication regimen. Additionally, they performed the same motor tasks as in the laboratory seven times a day at 30 min intervals. A total of seven time and frequency domain features were extracted from 5 s windows with 50% overlap. Then, a hierarchical Boosted Logistic Regression algorithm was used to estimate the tremor score for each limb during a motor task using the extracted features and ground truth labels. The actual and estimated tremor scores were averaged over all limbs and motor tasks for each repetition. The time that corresponds to the minimum tremor score was calculated in the laboratory and the home setting in order to identify the response to the medication intake. The root mean square error between the actual and predicted tremor scores was 0.21±0.14 using leave-one-out cross-validation. The mean values of the time that corresponds to the minimum tremor score were 92.8±23.5 min, 85.2±24.3 min, and 86.0±28.2 min using actual scores in the laboratory, estimated scores in the laboratory and estimated scores in the home environment, respectively. This study demonstrates that wearable sensors can be used to accurately predict the severity of tremor across different motor tasks. Furthermore, given that the temporal resolution of the monitoring was 30 min, the estimation results show that wearable sensors can effectively monitor the response of tremor to medication in the home setting.