Category: Technology
Objective: Use feature engineering and machine learning to evaluate digital biomarkers of early-stage Parkinson’s disease (PD) in the WATCH-PD study.
Background: WATCH-PD – a one-year longitudinal study – aims to relate remote sensor-based assessments to clinician-rated early-stage PD. Here, we used a combination of feature engineering and machine learning to evaluate the sensitivity of digital endpoints to detect early-stage PD status.
Method: 17 study sites enrolled PD (n=82) and healthy control (HC; n=50) participants into a one-year longitudinal study. BrainBaseline assessments of cognition, psychomotor performance, speech, and mobility were administered both on-site and at-home during the study. Continuous data were collected actively and passively on study provisioned Apple iPhones and Watches. Feature engineering routines estimated distributional properties of time- and frequency-dependent features derived from signal processing routines performed on continuous voice and accelerometry data sources. Machine learning algorithms were performed iteratively using Monte Carlo simulation (n=100). Each iteration randomly sorted features into independent training (90% of participants) and test sets. Feature selection was performed using linear regression to identify the most group-selective features. Logistic regression models of PD status were trained on independent features. Accuracy, sensitivity, and specificity were calculated from model predictions in the test set.
Results: At the time of analysis, participants completed 551 and 1,642 clinic and home sessions, respectively. Feature engineering yielded 3,622 features, 39.5% of which were selective for PD status. Features consistently selected across each Monte Carlo simulation (n=52) were associated with tremor-related activity during postural stability and at standing rest, wrist-to-trunk movement synchronization and tremor-related activity during active walking, and finger tapping efficiency. Model predictions yielded 85% accuracy, 83% sensitivity, and 86% specificity.
Conclusion: Remotely monitored sensor-based WATCH-PD assessments produced digital endpoints that predicted early-stage PD status with good accuracy, sensitivity, and specificity. Digital endpoints of primary interest were associated with tremor- and gait-related metrics. Further work is necessary to determine how well these digital endpoints track PD severity and progression.
To cite this abstract in AMA style:
D. Anderson, M. Merickel, B. Severson, D. Amato, T. Kangarloo, J. Edgerton, R. Dorsey, J. Adams, S. Jezewski, A. Keil, S. Johnson, M. Kantartjis, S. Polyak, J. Severson, J. Cosman. WATCH-PD: Detecting Early-Stage PD using Feature Engineering and Machine Learning in Remote Sensor-Based Assessments [abstract]. Mov Disord. 2022; 37 (suppl 2). https://www.mdsabstracts.org/abstract/watch-pd-detecting-early-stage-pd-using-feature-engineering-and-machine-learning-in-remote-sensor-based-assessments/. Accessed September 7, 2024.« Back to 2022 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/watch-pd-detecting-early-stage-pd-using-feature-engineering-and-machine-learning-in-remote-sensor-based-assessments/