Objective: To describe an approach using machine learning (ML) to develop composite digital measures that capture changes in physical activity in people with early-stage Parkinson’s disease (PD).
Background: Sensor-based digital measures capture real-world physical activity (e.g., daily step counts) and allow objective and continuous monitoring of physical function in people with early-stage PD [1, 2]. Composite measures, based on the combinations of single digital measures, may provide a more comprehensive understanding of one’s functioning while reducing noise. Yet, there is limited agreement on appropriate methods for developing composite measures such as selection of ML models, input features and approaches for model training.
Method: We included two years of passively collected data from the Personalized Parkinson Project via a wrist-worn sensor in 223 participants within ≤ 4 years of PD diagnosis at enrollment. Physical activity measures were derived from inertial motor unit (IMU) data [3, 4]. We trained ML models to generate composite measures using monthly aggregated single measures as input, with clinical assessment scores (MDS-UPDRS part 2 and 3) and patient reported outcomes (e.g., PDQ-39) as reference labels. Five ML models (linear regression with Lasso, ridge or elasticnet regularization, random forest regression, gradient boosted regression trees) were trained. We also compared different loss functions and evaluated model training approaches that do not require reference labels.
Results: The derived composite measures demonstrated statistically significant monotonic change over 2 years, with an absolute value of Cohen’s D ranging from 0.188 to 0.610 (random slope model with test for time effect, p < 0.05). As a comparison, the modified MDS-UPDRS Part 3 (excluding speech, facial expression and tremor items) has a Cohen’s D of 0.455. The test-retest reliability (ICC) of two consecutive months of the measures ranged from 0.705 to 0.911. The best measure had a Cohen’s D of 0.610, and ICC of 0.884.
Conclusion: This work provides a foundation to utilize ML techniques in building composite measures. The composite measures may provide a more comprehensive picture of one’s physical function and potentially be more sensitive to capture physical activity change along with the progression of PD.
References: [1] Chen et al. Parkinsonism Relat Disord. 2023
[2] Liu et al. Sci Transl Med. 2022
[3] Kowahl et al. JMIR Hum Factors. 2023
[4] Popham et al. JMIR Biomed Eng. 2023
To cite this abstract in AMA style:
K. Ho, C. Chen, S. Shin, S. Li, N. Kowahl, L. Evers, M. Meinders, L. Shih, B. Bloem, A. Siderowf, R. Kapur. Developing composite measures that track physical activity change in people with early-stage Parkinson’s disease using machine learning and wearable sensors [abstract]. Mov Disord. 2024; 39 (suppl 1). https://www.mdsabstracts.org/abstract/developing-composite-measures-that-track-physical-activity-change-in-people-with-early-stage-parkinsons-disease-using-machine-learning-and-wearable-sensors/. Accessed October 12, 2024.« Back to 2024 International Congress
MDS Abstracts - https://www.mdsabstracts.org/abstract/developing-composite-measures-that-track-physical-activity-change-in-people-with-early-stage-parkinsons-disease-using-machine-learning-and-wearable-sensors/