Clinical Validity of Wrist- and Trunk-Worn Sensor-Derived Data Models in Parkinson's Disease

Clinical rating scales for Parkinson's disease (PD) have limitations in the accurate assessment of disease severity, which may obscure treatment effects in clinical management and trials. Body-worn sensors can provide data for continuous and more precise monitoring of motor features of PD in patients' daily lives. However, little information exists on the clinical validity of sensor-derived data.

We assessed the clinical validity of outputs from three different machine learning models using trunk- or wrist-worn sensors in patients with PD, assessing their correlations with scores on clinical scales assessing motor severity and impact on function.

Wrist- and/or trunk-worn sensors were worn by patients with PD, who had been assessed using the MDS-UPDRS and the EQ-5D-5L, for up to one week. Output data were analyzed using three different algorithms: One trained on a publicly available dataset using trunk sensor data and two previously derived from wrist sensor data. Clinical validity was examined by examining correlations of sensor-derived outputs with individual items of the MDS-UPDRS and EQ-5D-5L.

For the trunk-worn sensor-derived outputs, the strongest positive correlations were found between output data and axial features such as arising from a chair, posture, and body bradykinesia and aspects of daily functioning on the MDS-UPDRS Part II and health-related quality of life (EQ-5D-5L domain) scores. For the wrist-worn sensor-derived outputs, the strongest positive correlations were seen between output data and postural tremor, rest tremor amplitude, and ability to undertake hobbies. Outputs from both body locations were correlated with MDS-UPDRS II and EQ-5D-5L scores (r > 0.7). In participants who wore both trunk and wrist sensors, percentage of time spent in different activities was similar between trunk- and wrist-worn devices, except for time spent "Lying down" derived from the trunk-worn sensor compared to time spent "Sleeping" derived from the wrist-worn sensor algorithm.

These data provide preliminary evidence for the clinical validity of single sensor assessments as measures of severity of motor features and motor functioning in patients with PD for use in clinical trials and practice. The results should be confirmed in large and more diverse populations and expanded to include other assessment methods such as laboratory-based motor measurements.