BACKGROUND
Smartphone-based digital biomarker (DB) assessments provide objective measures of daily-life tasks and thus hold the promise to improve diagnosis and monitoring of Parkinson's disease (PD). To date, little is known about which tasks perform best for these purposes and how different confounds including comorbidities, age and sex affect their accuracy.
OBJECTIVE
Here we systematically assess the ability of common self-administered smartphone-based tasks to differentiate PD patients and healthy controls (HC) with and without accounting for the above confounds.
METHODS
Using a large cohort of PD patients and healthy volunteers acquired in the mPower study, we extracted about 700 features commonly reported in previous PD studies for gait, balance, voice and tapping tasks. We perform a series of experiments systematically assessing the effects of age, sex and comorbidities on the accuracy of the above tasks for differentiation of PD patients and HC using several machine learning algorithms.
RESULTS
When accounting for age, sex and comorbidities, the highest balanced accuracy on hold-out data (67%) was achieved using relevance vector machine on tapping and when combining all tasks. Only moderate accuracies were achieved for other tasks (60% for balance, 56% for gait and 55% for voice data). Not accounting for the confounders consistently yielded higher accuracies of up to 73% (for tapping) for all tasks.
CONCLUSIONS
Our results demonstrate the importance of controlling DB data for age and comorbidities. They further point to a moderate power of commonly applied DB tasks to differentiate between PD and HC when conducted in poorly controlled self-administered settings.
Robust Detection of Parkinson's Disease Using Harvested Smartphone Voice Data: A Telemedicine Approach
