Analysis of movement patterns in individuals with Parkinson's disease experiencing motor fluctuations

Parkinson’s disease (PD) is a gradual neurodegenerative disorder defined by the loss of dopaminergic neurons and the presence of Lewy bodies in the basal ganglia area. The main manifestations of PD are motor symptoms, such as bradykinesia (i.e., slowness of movement), tremor, rigidity and balance instability, nonetheless nonmotor symptoms, as sleep disorders and dementia, are common too. Levodopa is the gold standard medication; its effectiveness is remarkable at the beginning, but it decreases with the progress of the therapy. Besides, motor complications appear as side-effects of the medication causing dyskinesia (i.e., involuntary movement) and motor fluctuations. To reduce the medication drawbacks and optimize the doses of levodopa, periodical assessments at the hospital are necessary. However, a longitudinal and continuous monitoring of individuals with late-stage PD would increase the efficacy of the drug titration. The use of wearable technologies can help the fulfillment of objective and longitudinal monitoring of motor symptoms and complications in unconstrained environments. This work aims to investigate the feasibility of a system able to track the severity of bradykinesia and motor fluctuations in naturalistic settings based on wearable sensors and machine learning (ML) techniques. The data set used for the study is part of the Blue Sky project and includes 25 participants with late-stage PD. The data are gathered during two visits, one in the laboratory and the other in a simulated apartment, using accelerometer sensors placed on the wrists and ankles. The tasks belong to standardized and activity of daily living (ADL) and the clinical scores follow the Unified Parkinson’s Disease Rating Scale (UPDRS). The data processing consists in filtering, resting period removal, signal segmentation, feature extraction, data cleaning, and training of the Random Forest (RF) algorithm. The signals are filtered to retain the frequency components related to bradykinesia. After the removal of the resting periods, the continuous signals are segmented into 5s windows to simplify the analysis. Once determined the feature set, the predictors are extracted from the windows and the redundant predictors are discarded applying a proposed method based on feature correlation and ReliefF. Four movement patterns groups are identified to reduce the complexity of the bradykinesia estimation. Data cleaning to discard outliers and improve class separation is implemented before the training of the RF regressor. The same pipeline is applied to the apartment data adding a movement pattern classifier before the model estimate. The cross-validation (CV) results on the laboratory data are obtained using k-fold and leave-one-subject-out (LOSO). The performance is measured in terms of root mean square error (RMSE) for regression tasks; whereas accuracy, specificity, and sensitivity are used for classification assignments. The validation results are encouraging, the overall RMSE is 0.5 in regression range between 0 and 3 with the LOSO CV, and the test results are promising for the longitudinal monitoring of bradykinesia and motor fluctuations in-home setting during ADL. In conclusion, this work demonstrates the feasibility of this approach applied in natural settings despite some limitations, such as the low number of subjects and a reduced amount of labels. This result can be a starting point for future improvements in PD severity monitoring.