JMIR Publications

ABSTRACT

Background:

Parkinson’s disease (PD) is a progressive neurological disorder with motor and non-motor symptoms that impair daily life. Conventional assessments such as the Unified Parkinson’s Disease Rating Scale (UPDRS) Hoehn & Yahr (H&Y) scales are subjective and limited to infrequent clinic visits, providing only brief snapshots of symptom severity. Wearable sensors combined with machine learning (ML) offer a means for continuous, objective symptom monitoring under real-world conditions. However, most existing approaches rely on multi-sensor systems that are impractical for long-term use.

Objective:

This paper describes the development and technical validation of ML models using single sensors to categorise activity measurements in people with PD in free-living conditions. Relationship between these activity measurements and clinical measurements is explored.

Methods:

ML models were trained for activity recognition on single sensor trunk-worn data from individuals from the general population. Signal statistical features, body orientation features, and Continuous wavelet transform (CWT) features were extracted, and Random Forest (RF) classifiers were trained to classify activities. The trained models were used to analyse sensor data collected from either trunk or wrist data in people with PD, and to predict activities and activity intensity. Correlation analysis is conducted between predicted activities features and clinical measurements.

Results:

Accuracy of activity prediction using trunk sensor was 93.4 (+/-8.6) on average and the F1-score for each activity was comparable to the literature. In the PD dataset, outputs included the percentage of time spent in different activities and statistical measures (mean, standard deviation, median, and variability) of activity and activity intensity, including for time spent with Activity, Walking, Standing and Lying, step count and duration, and Walking, Activities and heel strike acceleration. There were strong correlations of these outputs with corresponding clinical measurements. Significant correlations were observed between these sensor-derived activity features and clinical severity measures. Notably, the standard deviation and variability of step duration showed strong positive correlations (r > 0.81, p < 0.05) with motor severity subscores. In contrast, the median acceleration during walking and general activities correlated negatively with the H&Y stage (r = –0.64, p < 0.05).

Conclusions:

The findings suggest that a single-sensor approach holds promise for continuous PD monitoring and emphasis the potential of wearable sensor technology for continuous, objective, and cost-effective monitoring of PD progression. Clinical Trial: Antidepressants Trial in Parkinson’s Disease (ADepT-PD): ClinicalTrials.gov NCT03652870 Live Well with Parkinson’s (PD-Care): ISRCTN 92831552