Neural Network Models Predict Parkinson’s Disease Using Brief-action Motion Sensor Data
Suhrud Panchawagh1, Zion Zibly2, Veronica Santini3
1SKN Medical College, Pune, 2Department of Neurosurgery, 3Department of Neurology, Yale School of Medicine
Objective:
To develop and evaluate deep-learning models for predicting Parkinson’s disease using brief-action motion-sensor data from sit-to-stand and turning tasks.
Background:
Early initiation of lifestyle and pharmacologic treatments can enhance the quality of life, reduce costs, and improve outcomes in patients with Parkinson’s disease (PD). Currently, few biomarkers exist that can help with early diagnosis. Moreover, they are hindered by testing costs, invasiveness, and lead time. Motion sensor data effectively captures spatial movement patterns, which can serve as a biomarker collected in the clinic, community, or resource-limited and rural settings where medical practice is dictated primarily by generalists. We propose a deep-learning model using brief action motion-sensor data to predict PD.
Design/Methods:
We utilized data from the REMAP open dataset, which includes 403 sit-to-stand (STS) and 1749 turning episodes. We built neural-network models using STS, 2D, and 3D turning data and an ensemble model using STS and 3D turning data. SHAP (SHapley Additive exPlanations) values were calculated to identify key features influencing predictions.
Results:
The median durations of the STS and turning episodes were 1.95 and 1.74 seconds. The STS model demonstrated an accuracy (±standard error) of 77% ± 1.3%, precision of 74.7% ± 1.9%, sensitivity of 80.5% ± 1.8%, and specificity of 73.6% ± 1.9%. The 2D and 3D turning models and the ensemble STS and 3D turning models had accuracies of 69.2%, 72%, and 76.6%, respectively. SHAP analysis revealed that the whole episode duration was the most influential feature in the STS model, while the number of turning steps for the turning models consistently explained predictions.
Conclusions:
Our models using brief-action motion-sensor data performed well in detecting PD during STS and turning tasks. Using multicentric data to improve generalizability, these models have the potential for deployment in resource-limited settings, enhancing personalized medicine and improving quality of life through earlier referrals and treatment initiation.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.