Predicting the severity of Parkinson’s disease symptoms with smartwatches during daily living

Abstract

Parkinson’s Disease (PD) is a neurodegenerative disease that affects millions of people worldwide and can significantly affect the quality of life. Currently, the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) test evaluates the disease’s progression and is considered the gold standard. However, the test is done on average once or twice a year to assess motor and non-motor symptoms related to PD. It only provides a snapshot of the symptoms on a given day which is a significant limitation of the test. There is a need for passive, longitudinal, and in-the-wild home monitoring of the disease as PD symptoms can fluctuate for various reasons, such as the quality of sleep, from one day to the next. Monitoring the symptoms during daily life can provide essential insights and lead to better health care decisions for patients by providing information about how severe the symptoms are on average throughout the whole year. As a result, clinicians could find the most optimal medication schedule and dosage for patients, which has the potential to reduce side-effects of the medication like dyskinesia. Most of the work has been done either in clinical environments or at home with scripted tasks that subjects need to complete at specific times, which can be an additional burden. This is why this work aims to quantify the medication status (on/off), the severity of tremor, and dyskinesia from a single wrist-worn smartwatch during passive monitoring. Using the accelerometer of the Apple Watch, we propose three approaches to solve this problem. The first one is based on time series features extraction with an Extreme Gradient Boosting (XGBoost), while the second uses embeddings with different classifiers like Probabilistic Linear Discriminant Analysis (PLDA), K-Nearest Neighbors (KNN), and Support Vector Regression (SVR). The third approach is a fusion of both approaches using a simple average. We experimented with different data pre-processing methods, such as using a High-Pass Filter (HPF) to remove the gravity component. We also detected and removed inactivity in the signals. Furthermore, we artificially generated new samples with various data augmentation techniques such as linear combination, Gaussian noise, resampling, and rotation. Finally, the third approach was the most successful and achieved a weighted Mean Square Error (MSE) of 1.129, 0.429, and 0.462 for on/off, tremor, and dyskinesia. Thus, for each symptom, a different combination of data pre-processing and augmentation successfully improved the overall MSE.

Date	1 Jun 2021
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Patrick Cardinal (Supervisor) & Laureano Moro-Velazquez (Co-supervisor)