[Paper Review] A Multi-layer Gaussian Process for Motor Symptom Estimation in People with Parkinson's Disease

M. Lang, Franz Pfister+8 more|arXiv (Cornell University)|Aug 31, 2018

Parkinson's Disease Mechanisms and Treatments24 references3 citations

TL;DR

This paper proposes a multi-layer Gaussian process model that estimates the severity of Parkinson's disease motor symptoms—tremor, non-tremulous bradykinesia, and non-tremulous dyskinesia—using inertial sensor data from wrist-worn wearables during unscripted daily activities. By hierarchically stacking Gaussian process models trained on clinically rated MDS-UPDRS annotations, the method achieves robust, objective, and continuous symptom monitoring in real-world settings with high agreement to clinical assessments.

ABSTRACT

The assessment of Parkinson's disease (PD) poses a significant challenge as it is influenced by various factors which lead to a complex and fluctuating symptom manifestation. Thus, a frequent and objective PD assessment is highly valuable for effective health management of people with Parkinson's disease (PwP). Here, we propose a method for monitoring PwP by stochastically modeling the relationships between their wrist movements during unscripted daily activities and corresponding annotations about clinical displays of movement abnormalities. We approach the estimation of PD motor signs by independently modeling and hierarchically stacking Gaussian process models for three classes of commonly observed movement abnormalities in PwP including tremor, (non-tremulous) bradykinesia, and (non-tremulous) dyskinesia. We use clinically adopted severity measures as annotations for training the models, thus allowing our multi-layer Gaussian process prediction models to estimate not only their presence but also their severities. The experimental validation of our approach demonstrates strong agreement of the model predictions with these PD annotations. Our results show the proposed method produces promising results in objective monitoring of movement abnormalities of PD in the presence of arbitrary and unknown voluntary motions, and makes an important step towards continuous monitoring of PD in the home environment.

Motivation & Objective

• To enable continuous, objective monitoring of Parkinson’s disease motor symptoms in unstructured daily environments.
• To address the challenge of distinguishing PD-related movement abnormalities from arbitrary voluntary motions in free-living conditions.
• To estimate not only the presence but also the severity of tremor, bradykinesia, and dyskinesia using clinical severity ratings as supervision.
• To develop a hierarchical modeling framework that captures the distinct characteristics of different PD symptom classes.
• To validate the method’s reliability in real-world settings using wearable sensor data and clinician-annotated severity scores.

Proposed method

• The method employs a multi-layer Gaussian process (GP) architecture, with each layer independently modeling one of three PD symptom classes: tremor, non-tremulous bradykinesia, and non-tremulous dyskinesia.
• Each GP model is trained using input features derived from wrist-worn inertial sensor data, including time-domain and time-frequency domain representations such as spectrograms and wavelet coefficients.
• The models use clinical severity scores from the MDS-UPDRS as scalar outputs, enabling regression of symptom severity on a 0–4 scale.
• The hierarchical stacking allows for independent modeling of symptom-specific dynamics while preserving interpretability and robustness to unknown voluntary movements.
• A squared exponential kernel is used for each GP, with hyperparameters optimized via marginal likelihood maximization to capture non-linear mappings from sensor inputs to symptom severity.
• The final prediction for each symptom is a Gaussian distribution, providing both a severity estimate and a confidence interval.

Experimental results

Research questions

• RQ1 Can a multi-layer Gaussian process model accurately estimate the severity of multiple Parkinson’s disease motor symptoms from unstructured wrist motion data?
• RQ2 How well can the model disambiguate PD-related movement abnormalities from arbitrary voluntary movements in free-living conditions?
• RQ3 To what extent does the hierarchical GP structure improve the estimation of distinct symptom classes compared to single-model approaches?
• RQ4 Can the model achieve high agreement with clinical annotations despite the absence of task-specific instructions or controlled activities?
• RQ5 Does incorporating time-frequency domain features enhance the model’s ability to detect subtle or non-tremulous motor symptoms?

Key findings

• The proposed multi-layer GP model achieved strong agreement between predicted symptom severities and clinician-annotated MDS-UPDRS scores, demonstrating high reliability in real-world settings.
• The model successfully estimated the severity of tremor, non-tremulous bradykinesia, and non-tremulous dyskinesia without requiring patients to perform specific tasks.
• The use of time-frequency domain features significantly improved the detection of non-tremulous motor abnormalities, which are often challenging to capture with time-domain features alone.
• The hierarchical structure enabled robust estimation even in the presence of unknown and arbitrary voluntary movements, maintaining high predictive accuracy.
• The model's predictive confidence intervals were well-calibrated, indicating reliable uncertainty estimates for clinical decision support.
• The results suggest that the method is a viable step toward continuous, objective, and home-based monitoring of Parkinson’s disease progression.