Continuous and Unconstrained Tremor Monitoring in Parkinson's Disease Using Supervised Machine Learning and Wearable Sensors

Abstract

Background: Accurately assessing the severity and frequency of fluctuating motor symptoms is important at all stages of Parkinson's disease management. Contrarily to time-consuming clinical testing or patient self-reporting with uncertain reliability, recordings with wearable sensors show promise as a tool for continuously and objectively assessing PD symptoms. While wearables-based clinical assessments during standardised and scripted tasks have been successfully implemented, assessments during unconstrained activity remain a challenge.

Methods: We developed and implemented a supervised machine learning algorithm, trained and tested on tremor scores. We evaluated the algorithm on a 67-hour database comprising sensor data and clinical tremor scores for 24 Parkinson patients at four extremities for periods of about 3 hours. A random 25% subset of the labelled samples was used as test data, the remainder as training data. Based on features extracted from the sensor data, a Support Vector Machine was trained to predict tremor severity. Due to the inherent imbalance in tremor scores, we applied dataset rebalancing techniques.

Results: Our classifier demonstrated robust performance in detecting tremor events with a sensitivity of 0.90 on the test-portion of the resampled dataset. The overall classification accuracy was high at 0.88.

Conclusion: We implemented an accurate classifier for tremor monitoring in free-living environments that can be trained even with modestly sized and imbalanced datasets. This advancement offers significant clinical value in continuously monitoring Parkinson's disease symptoms beyond the hospital setting, paving the way for personalized management of PD, timely therapeutic adjustments, and improved patient quality of life.

Figure 1

Sporadic clinical encounters do not provide an accurate portrayal of the disease course. Important bouts of severe symptoms, if they do not happen during clinical visits, might be missed. This leads to a gap between clinical assessment and patient self-report, which is the currently available system used to monitor the disease severity between appointments. This self-report, oftentimes in the form of diary entries, can however lack in precision, objectivity, and compliance on the side of the patient. Continuous monitoring would allow for automated and objective observation of disease state and progression with only moderate effort on the side of both clinicians and patients.

Figure 2

Depiction of various activities performed by patients during tremor assessment sessions. Clinicians conducted evaluations at three-minute intervals in an unconstrained environment simulating daily living scenarios, including reading, device usage, and light physical tasks.

Figure 3

Preprocessed IMU data capturing tremor dynamics. The composite accelerometer signal, derived from the vector sum of three-axis data, is shown at the top, indicating fluctuating tremor magnitudes. Initially, the tremor intensity sharply peaks, diminishes briefly, rises again, and tapers off. Below, the triaxial gyroscope data reveal that rotational motion, particularly along the X-axis, is not consistently synchronized with these variations in tremor magnitude. The initial strong rotational activity is not mirrored during subsequent peaks of translational tremor, illustrating the distinct behaviours of these tremor components.

Figure 4

The dataset is highly imbalanced, with Score 0 (no tremor) representing 91.5% of all observations. Score 1 (mild tremor) represented only 7.5% of all observations and Score 2 (strong tremor) only 1%.

Figure 5

Balanced dataset after the resampling strategy. The diagonally striped, blue portion in Score 2 represents synthetic data.

Figure 6

2D principal component analysis of (a) the initial dataset and (b) the final dataset after applying the resampling strategy. While in (a) the tremor scores are highly imbalanced, the proportions of the three classes have been equalised in (b). Synthetic data are represented with a “+” sign. PCA embedding is based on the 30 best-performing features in the initial dataset.

Figure 7

(a) Confusion matrix and (b) normalised confusion matrix for the test part of the in-sample dataset. Consistent with the principal component visualisations, there is a degree of overlap between the classes that results in some instances of adjacent classes being misclassified. The overall classification accuracy is 0.88.

Figure 8

(a) Confusion matrix and (b) normalised confusion matrix for the out-of-sample predictions on the instances discarded during the undersampling technique. With an overall classification accuracy of 0.94, out-of-sample performance is comparable to in-sample performance.