Authors trained a convolutional neural network on frequency features from a single lumbar accelerometer during quiet standing and achieved ~98–99% accuracy distinguishing 40 early untreated PD patients from 79 controls.

Provides a simple, noninvasive, high-accuracy candidate biomarker for detecting subtle early postural impairment that could aid earlier diagnosis, trial enrollment, and objective outcome measurement, though it does not advance mechanistic or therapeutic targets.

Despite evidence of early neurodegeneration, postural instability is commonly associated with later stages of Parkinson's disease (PD), mainly due to a lack of sensitive measures. Here, we aim to provide a sensitive, easily obtainable objective measure of postural instability for earlier clinical detection. We assessed postural sway in 40 newly diagnosed, untreated individuals with PD and 79 age-matched healthy controls while they stood quietly for 30 seconds with their eyes open and feet together. Body sway was recorded with a single accelerometer placed at the lumbar spine. We trained a convolutional neural network (CNN) to distinguish between the groups based on the frequency information of their sway signals. Our models reached an average accuracy, sensitivity, and specificity of 98.9%, 97.7%, and 98.9%, respectively. This suggests that characteristic frequency features of postural sway reflect subtle postural impairments in early PD, with great potential to translate into clinical applications.