F62. Automatic detection of ON/OFF states in Parkinson disease patients using wearable inertial sensors. (May 2018)

• Record Type: Journal Article
• Authors: Anand, Vibha; Bilal, Erhan; Ramos, Vesper; Naylor, Melissa; Demanuele, Charmaine; Zhang, Hao; Amato, Stephen; Wacnik, Paul; Hameed, Farhan; Kangarloo, Tairmae; Ho, Bryan; Erb, Kelley; Karlin, Dan
• Is Part Of: Clinical neurophysiology. Volume 129(2018)Supplement 1
• Page Start: e90
• Publication Date: 2018-05
• Subjects: Neurophysiology -- Periodicals; Electroencephalography -- Periodicals; Electromyography -- Periodicals; Neurology -- Periodicals; 612.8
• DOI: 10.1016/j.clinph.2018.04.225
• Languages: English
• ISSNs: 1388-2457

Abstract

Abstract : Introduction: Reliably detecting ON/OFF states is important for monitoring PD treatment and progression. Currently, subjective patient diaries capture this information. We investigate if detection of motor signs of ON/OFF states can be achieved by using kinematic measurements from wearable sensor technology combined with a machine learning (ML) pipeline. Methods: Twenty-five PD subjects (19 males, 69 ± 7 years) taking levodopa performed 10-m Instrumented Stand and Walk (ISAW) tests in their ON and OFF states while wearing Ambulatory Parkinson Disease Monitoring (APDM) sensors on their sternum, wrists, lumbar and lower extremities. A neurologist scored each ISAW according to the MDS-UPDRS-III. We analyzed 98 kinematic features for significance to neurologist total motor score and ON/OFF using both statistical (repeated-measures ANOVA, step-wise mixed-model regression, likelihood-ratio test, ridge regression) and ML methods. Results: Twenty-two features significantly differed between patient reported ON/OFF states, with the most significant being trunk transverse range-of-motion (RofM), arm RofM, mid-swing elevation, stride length, turn velocity, steps in turn and toe out angles. Estimates from regression model showed average difference of 14 points between OFF/ON states in total UPDRS score and 9 points when adjusted for 5 significant features for individual baseline (mean trunk transverse RofM, right arm RofM, and toe out angle having highest effect; coeff. −8.67, −5.25, −3.36 respectively). Several approaches were employed for predicting ON/OFF states based on these features: direct binary classification (acc = 0.56), regression to total UPDRS score (acc = 0.76), regression to PIGD sub-score (acc = 0.64), and classification of ON–OFF/OFF–ON transitions using feature differences (Naïve Bayes: acc = 0.74, AUC = 0.78; Random Forest: acc = 0.76, AUC = 0.90). Conclusion: Wearable inertial sensors hold promise for detecting ON/OFF states in PD patients using an augmented ML approach. This could be particularly useful for monitoring response to therapy in an outpatient setting.