Passive Monitoring of Parkinson Tremor in Daily Life: A Prototypical Network Approach

Luc J. W. Evers, Yordan P. Raykov, Tom M. Heskes, Jesse H. Krijthe, Bastiaan R. Bloem, Max A. LittleSensors 2025 · OpenReview

TL;DR

This paper introduces a novel "prototypical network" approach for passively monitoring Parkinson's disease (PD) tremor using wrist-worn accelerometers. By leveraging domain knowledge to define "prototypical" examples of tremor and non-tremor activities, their method achieves robust tremor detection in real-world settings, even with limited labeled data, outperforming traditional machine learning models.

What Problem Does This Paper Address?

Parkinson's disease (PD) significantly impacts millions worldwide, with tremor being a key symptom. Current clinical assessments of tremor are often snapshots, failing to capture the day-to-day variability influenced by medication, stress, and the patient's environment. This makes it challenging to accurately track disease progression and treatment effectiveness. Wearable sensors offer a promising solution for continuous, passive monitoring, but developing reliable tremor detection models for real-life conditions is difficult due to the inherent heterogeneity of both tremor (different types and severities) and non-tremor activities (a wide range of daily behaviors). Existing methods often struggle with this variability, especially when labeled data is scarce, which is common in studies of unscripted daily life.

Methods

The core innovation is the use of prototypical networks, a type of neural network that incorporates domain expertise through "prototypical examples." Instead of learning solely from raw data, these networks learn by comparing input data to predefined, representative examples of different classes.

Here's a breakdown:

  • Domain Expertise: The researchers, in consultation with a neurologist, defined specific "prototypical" examples for different tremor sub-classes (e.g., wrist flexion-extension with arm supported vs. free) and non-tremor activities (e.g., walking, sitting, driving). These examples capture the real-world variability.
  • Prototypical Networks: They developed both single-layer and a novel two-layer prototypical Radial Basis Function (RBF) network.
  • Single-layer: Directly maps input features to class probabilities using RBF kernels centered around the prototypical examples.
  • Two-layer: Introduces a dimensionality reduction layer before the classification layer. This allows the network to learn a more compact representation of the data before comparing it to prototypes, potentially improving robustness and handling complex feature spaces.
  • Training: The training process involves inferring basis parameters from the prototypical examples and then fitting a logistic model. A key aspect is their custom training procedure for the two-layer network, which optimizes intermediate embeddings to mitigate vanishing gradient issues common in deep networks.

Datasets

The study utilized data from the Parkinson@Home Validation Study. This dataset is notable because it includes:

  • Participants: 8 PD patients with tremor, 16 PD patients without tremor, and 24 age-matched controls.
  • Data Collection: Wrist-worn accelerometer data collected during unscripted daily life activities in the participants' own homes.
  • Ground Truth: Synchronous expert video annotations for the presence and severity of tremor, captured during these activities. This is a significant advantage over studies relying on infrequent clinical assessments.

The researchers also curated annotated prototypical examples for various tremor sub-classes and non-tremor activities, derived from the video recordings, to inform their model training.

Key Insights and Ideas

  • Domain Knowledge is Powerful: Explicitly incorporating clinical expertise through prototypical examples significantly enhances model performance, especially in complex, real-world scenarios where data is limited.
  • Heterogeneity is Key: The approach directly addresses the challenge of diverse tremor types and daily activities by defining distinct sub-classes, leading to more robust detection.
  • Two-Layer Networks Offer Advantages: The proposed two-layer prototypical network demonstrated improved sensitivity and reduced variability compared to the single-layer version and baseline methods, suggesting that learning a compressed representation before classification is beneficial.
  • Data Scarcity is Manageable: The learning curves indicate that prototypical networks can achieve acceptable performance with relatively small amounts of labeled training data, a crucial advantage for passive monitoring applications.

Key Results

  • Overall Performance: The two-layer prototypical network achieved the highest average sensitivity (66%) and AUROC (87%) among the tested methods, outperforming logistic regression and random forest classifiers.
  • Robustness:
    • Activities: The prototype models showed higher specificity during rarer non-tremor activities (e.g., "Driving bike/car," "Suspicious activities") with lower patient-to-patient variance.
  • Tremor Sub-classes: The prototype models demonstrated higher sensitivity for most tremor sub-classes, particularly "Tremor during Gait," where they significantly outperformed baseline models.
  • Learning Curves: Prototypical models showed higher AUROC with limited training data (2 minutes or less) compared to baseline models.
  • Patient-Specific Performance: The two-layer model showed good individual performance, with AUROCs generally above 0.80, and sensitivity correlating with rest tremor severity.

Limitations and Future Work

  • Limited Sample Size: The study included only 8 PD patients with tremor, making it difficult to conduct rigorous statistical comparisons and limiting the generalizability of the findings. Future work should validate these results on larger, independent datasets.
  • Prototypical Example Selection: The current selection of prototypical examples is a starting point. Future work could expand this to include more detailed tremor sub-types (e.g., rest, re-emergent rest, action tremor) and potentially quantify their association.
  • Time-Dependent Nature: The current models do not fully exploit the temporal dynamics of tremor or human behavior. Future research could integrate these models into sequential models like Hidden Markov Models (HMMs) to improve detection and infer sequential predictors of tremor.
  • Clinical Utility: While the models show promise for passive monitoring, their added value for clinical decision-making, quality of life, and patient functioning needs further elucidation.

Summary updated

Abstract

Objective and continuous monitoring of Parkinson’s disease (PD) tremor in free-living conditions could benefit both individual patient care and clinical trials, by overcoming the snapshot nature of clinical assessments. To enable robust detection of tremor in the context of limited amounts of labeled training data, we propose to use prototypical networks, which can embed domain expertise about the heterogeneous tremor and non-tremor sub-classes. We evaluated our approach using data from the Parkinson@Home Validation study, including 8 PD patients with tremor, 16 PD patients without tremor, and 24 age-matched controls. We used wrist accelerometer data and synchronous expert video annotations for the presence of tremor, captured during unscripted daily life activities in and around the participants’ own homes. Based on leave-one-subject-out cross-validation, we demonstrate the ability of prototypical networks to capture free-living tremor episodes. Specifically, we demonstrate that prototypical networks can be used to enforce robust performance across domain-informed sub-classes, including different tremor phenotypes and daily life activities.
BibTeX
@inproceedings{evers2025passive,
  author = {Luc J. W. Evers and Yordan P. Raykov and Tom M. Heskes and Jesse H. Krijthe and Bastiaan R. Bloem and Max A. Little},
  title = {{Passive Monitoring of Parkinson Tremor in Daily Life: A Prototypical Network Approach}},
  booktitle = {Sensors},
  year = {2025},
  url = {https://openreview.net/forum?id=KCoUfJ7I0j},
}
Figure