Session Information
Date: Thursday, June 8, 2017
Session Title: Parkinson’s Disease: Clinical Trials, Pharmacology And Treatment
Session Time: 1:15pm-2:45pm
Location: Exhibit Hall C
Objective: We describe relevant aspects and data processing steps for accelerometry and gyroscopy recordings made with a wrist-worn low-cost sensor device. In order to generalize and compare results between subjects, we propose the use of a body-centered frame of reference for motion data.
Background: Body-worn sensor data can assess physiological motion as well as movement disorders, e.g. Parkinson’s Disease (PD) with high temporal resolution in free-living situations. Appropriate evaluation of this data opens opportunities for objective symptom measurement. The collection of large amounts of data by use of commercially available sensors poses hitherto unknown challenges to data validity.
Methods: Ethical approval was obtained from Technical Uni. of Munich. For this principal of proof study, we included 27 PD patients and 8 healthy controls. Raw data was recorded at a rate of 62.5 Hz using the MS band 2 (Microsoft). We applied low-pass filters to reduce accel. noise, and high-pass filters to reduce impact of gyro. integration drift. Data was normalized by rotating it from the sensor coordinate system to the reference coordinate system. Sensor pose was estimated using quaternion based complementary filters. Ground truth data was created to validate such pose estimation (PE). Movements reflecting gradings of severity of MDS-UPDRS items, e.g. bradykinesia, were recorded using the MS band 2 and synchronized to data captured by a motion capture system (Qualisys Track Manager, QTM). The QTM pose was compared to the PE from the wristband data. The data was used to train neural networks (NN) as described in Pfister et al.
Results: Specific steps of data processing were required to reduce inter-subject variability caused by differences in sensor orientation and placement. The PE obtained by integration from the sensor signals accurately reflects the motion patterns, visible in the corresponding QTM recordings, and approves our steps of processing. Preprocessing increased the precision of NN to correctly classify short data segments as clinical meaningful data by 55.5%.
Conclusions: Wrist-worn sensors can be used to describe spontaneous movement with unprecedented reliability. However, a high degree of sophistication and preprocessing is necessary to overcome problems inherent to this data type. The normalization of the coordinate system can be used to improve inter-subject recognition of NN.
References: Pfister, et al., Deep Learning in objective classification of spontaneous movement of patients with Parkinson’s Disease using large-scale free-living sensor data, submitted abstract for the Movement Disorder Society 21st International Congress, Vancouver (June 2017)
To cite this abstract in AMA style:
D. Pichler, M. Lang, D. Kulić, F. Pfister, G. König, T. Um, A. Ahmadi, S. Endo, F. Achilles, K. Abedinpour, K. Bötzel, A. Ceballos-Baumann, S. Hirche, U. Fietzek. Acquisition, Validation and Preprocessing of Wrist-Worn Sensor Data in Patients with Parkinson’s Disease and Healthy Controls [abstract]. Mov Disord. 2017; 32 (suppl 2). https://www.mdsabstracts.org/abstract/acquisition-validation-and-preprocessing-of-wrist-worn-sensor-data-in-patients-with-parkinsons-disease-and-healthy-controls/. Accessed December 10, 2024.« Back to 2017 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/acquisition-validation-and-preprocessing-of-wrist-worn-sensor-data-in-patients-with-parkinsons-disease-and-healthy-controls/