Objective: To identify features from wearable sensors that can predict disease severity as measured by the MDS-UPDRS.

Background: Wearable sensors can provide quantitative assessments of gait and posture in patients with Parkinson’s disease (PD). Measurements from these sensors enable a more granular assessment of disease symptoms and can be related to MDS-UPDRS sub-scores; however, the reliability of these relationships is not clear [1]. Machine learning has been applied to wearable sensor features to identify and assess PD severity [1].

Method: 164 PD patients (age: 64.6±10.1, male: 113, H&Y: 1.7±0.59, MDS-UPDRS III: 22.5±11.2 during ON state) completed an in-clinic 1-minute walking test at their self-selected pace and postural sway task (standing stationary with feet together for 30 seconds with eyes open) while wearing 5 sensors (ADPM, Inc) containing a 3D accelerometer and 3D gyroscope on their wrists, shanks, and lumbar spine [2]. 22 feature selection algorithms were used to identify a subset of gait and posture features (from a total of 47 sway and 59 gait features) that best predicted selected relevant MDS-UPDRS sub-scores as a reference standard for PD severity (3.10 Gait, 3.11 Freezing of Gait; 3.12 Postural Stability, 3.13 Posture). 10-fold cross-validation was performed for each algorithm using a Random Forest classifier compared to the selected MDS-UPDRS sub-scores to identify a set of most relevant features. Each feature set was then tested with 25 machine learning classification models. Feature sets with at least 75% classification accuracy after a 10-fold cross validation were included in the analysis.

Results: When comparing walking to MDS-UPDRS subscore 3.11, stride length was the most commonly selected feature, followed by gait speed (both average and coefficient of variation), and cadence. For postural sway compared to 3.12, anterior-posterior jerk and total normalized jerk were the most common features, followed by total sway area, and then average anterior-posterior velocity and path. Sub-scores 3.10 and 3.13 did not produce any feature sets above 75% accuracy.

Conclusion: Features extracted from wearable sensors during walking and postural sway were sensitive to PD severity as measured by MDS-UPDRS. The findings highlight the potential of including quantitative gait and postural control assessments using wearable sensors in future clinical trials to make predictions of a patient’s PD severity.

References: [1] Ramdhani RA, Khojandi A, Shylo O, Kopell BH. Optimizing Clinical Assessments in Parkinson’s Disease Through the Use of Wearable Sensors and Data Driven Modeling. Front Comput Neurosci. 2018; 12:72. [2] Mancini M, Horak FB. Potential of APDM mobility lab for the monitoring of the progression of Parkinson’s disease. Expert Rev Med Devices. 2016; 13: 455-62.

« Back to 2019 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/identifying-in-clinic-wearable-sensor-features-that-predict-parkinsons-disease-severity/