Quantifying device type and handedness biases in a remote Parkinson's disease AI-powered assessment.

We investigate issues pertaining to algorithmic fairness and digital health equity within the context of using machine learning to predict Parkinson's Disease (PD) with data recorded from structured assessments of finger and hand movements. We evaluate the impact of demographic bias and bias related to device type and handedness. We collected data from 251 participants (99 with PD or suspected PD, 152 without PD or any suspicion of PD). Using a random forest model, we observe 92% accuracy, 94% AUROC, 86% sensitivity, 92% specificity, and 84% F1-score. When examining only F1-score differences across groups, no significant bias appears. However, a closer look reveals bias regarding positive prediction and error rates. While we find that sex and ethnicity have no statistically significant impact on PD predictions, biases exist regarding device type and dominant hand, as evidenced by disparate impact and equalized odds. Our findings suggest that remote digital health diagnostics may exhibit underrecognized biases related to handedness and device characteristics, the latter of which can act as a proxy for socioeconomic factors.