Comprehensive Summary
This paper investigates whether artificial intelligence can fairly predict Parkinson’s disease (PD) through remote motor and memory tasks. The authors developed a web based tool where participants used their keyboard and mouse to complete tracing, pressing, and memory exercises. Data from 251 individuals, including those with PD or suspected PD and healthy controls, were analyzed using multiple machine learning models. The models reached strong performance overall, with accuracy above 90%. While sex and race did not significantly affect outcomes, the results revealed differences tied to the device used (Mac versus Windows) and whether participants were left or right handed. These technical and physical factors influenced the likelihood of receiving a PD prediction, showing that even well performing models may contain hidden biases.
Outcomes and Implications
The study highlights how remote digital health tools can introduce inequities if subgroup differences are overlooked. Device type may reflect underlying socioeconomic status, and the findings show that left handed users and Windows users were less likely to be flagged for PD. This raises concerns that AI systems could unintentionally disadvantage certain groups, even as they aim to expand access to neurological screening. For these tools to move closer to clinical use, larger and more balanced datasets are needed, along with careful design that accounts for hardware and user diversity. Without such safeguards, digital diagnostics risk reinforcing disparities in care rather than reducing them.