The late stage of Parkinson’s disease (PD) is accompanied by unpredictable motor fluctuations which cannot be treated sufficiently with a medication schedule with fixed time intervals. At this point, feedback-based medication timing based on the medication state would satisfy the
...
The late stage of Parkinson’s disease (PD) is accompanied by unpredictable motor fluctuations which cannot be treated sufficiently with a medication schedule with fixed time intervals. At this point, feedback-based medication timing based on the medication state would satisfy the needs of persons with PD. Wearable sensors offer the ability to objectively and continuously monitor PD-related features in an unsupervised way during daily life. Sensor fusion of a shank-worn accelerometer (Cue2walk wearable device) and a heart rate sensor enables comprehensive state monitoring through the integration of physical and physiological parameters. This study aimed to identify features, retrieved from the Cue2walk wearable device and/or a heart rate sensor, correlated with the medication state in PD, in an unsupervised way during daily life. Triaxial accelerometer and heart rate data were collected from nine persons with PD for seven days. Features in the time and frequency domain were extracted and a value was assigned to each stride. The frequency features were calculated by including walking bouts of at least 10 s. The feature values were averaged over seven block sizes. The medication state was modelled by assuming a sinusoidal form and was based on the medication intake schedule. For each averaging block size, the correlation between the features and the medication state was calculated. Due to a hardware problem, only data from one participant could be included. Stride time variability and displacement, velocity, acceleration, and covariance features resulted in the largest significant correlation with the medication state in this participant. The reliability of the correlations was limited by the short walking bout duration in the calculation of the frequency features and by the simplified model of the medication state. Therefore, future research should collect data on a longer term and include walking bouts of 25 s in the calculation of the frequency features. Moreover, medication intake should be logged by the participants. Consequently, personalised digital biomarkers for medication state monitoring can be developed in an unsupervised way, by taking into account multicollinearity.
