Parkinson's Disease Tremor Detection in the Wild Using Wearable Accelerometers

Rubén San Segundo, Ada Zhang, Alexander Cebulla, Stanislav Panev, Griffin Tabor, Katelyn Stebbins, Robyn E. Massa, Andrew S. Whitford, Fernando De la Torre, Jessica K. HodginsSensors 2020 · DOI · OpenReview

TL;DR

This paper introduces a novel system using wrist-worn accelerometers to detect Parkinson's Disease (PD) tremor in real-world, uncontrolled environments. By combining a specialized tremor spectrum extraction technique with a deep neural network, the system achieves high accuracy and shows promise for continuous, objective PD symptom monitoring.

What Problem Does This Paper Address?

Parkinson's Disease (PD) is a chronic disorder characterized by motor symptoms, primarily tremor. Current methods for assessing PD severity, like the Unified Parkinson's Disease Rating Scale (UPDRS), are subjective and infrequent. While patient diaries offer more frequent tracking, they are burdensome and prone to recall bias. The ideal solution is an autonomous, continuous monitoring system that can accurately detect PD symptoms in everyday life, providing objective data for disease progression assessment and treatment adjustment. However, detecting PD tremor in "the wild" (uncontrolled environments) is challenging due to variability in patient symptoms, daily activities, and the difficulty of obtaining precise labels for training and validation.

Methods

The researchers explored various feature extraction techniques and classification algorithms. A key innovation is a tremor spectrum extraction method based on Non-Negative Matrix Factorization (NMF). This unsupervised pre-processing step aims to isolate the tremor signal from the overall accelerometer data before feature extraction. They then evaluated several feature sets:

  • Energy in the 3-9 Hz Band: A simple benchmark.
  • Welch's Power Spectral Density (PSD): A standard spectral analysis technique.
  • Common Baseline Features: A comprehensive set of time and frequency domain features.
  • Mel Frequency Cepstral Coefficients (MFCCs): Features commonly used in speech processing, adapted for motion data.
  • MFCCs-T/NT: MFCCs applied after the tremor spectrum extraction.
  • CNN (Convolutional Neural Network): A deep learning model trained directly on raw accelerometer data.
  • CNN-T/NT: A CNN trained on the spectra obtained after the tremor spectrum extraction.

For classification, they compared Random Forests (RF) and Multi-Layer Perceptrons (MLP). The best-performing system, CNN-T/NT + MLP, was then compared against previous state-of-the-art methods. They also investigated the system's ability to predict patient self-reported tremor levels and proposed a new metric: the percentage of tremor time over longer intervals, which is easier to estimate than precise tremor event detection.

Datasets

The study utilized two main datasets:

  • Laboratory Recordings (LAB): Data collected from 12 PD patients in a controlled lab setting, performing daily living activities and UPDRS tasks. This dataset provided accurate, manually annotated ground-truth labels for tremor presence and absence.
  • In-the-Wild Recordings (WILD): Data collected from 6 PD patients wearing accelerometers for four weeks in their natural environments. Participants provided weak labels (Almost none, Half the time, Almost always) via a smartphone app, indicating the approximate amount of tremor experienced in the preceding five minutes.

Key Insights and Ideas

  • Tremor Spectrum Extraction is Crucial: The novel tremor spectrum extraction pre-processing step significantly boosted the performance of both MFCCs and CNN features. This suggests that isolating the tremor signal before feature extraction is more effective than using raw data or standard spectral analysis alone.
  • Deep Learning Outperforms Handcrafted Features (with enough data): The CNN-T/NT approach, which learns features directly from the processed spectra, achieved the highest performance. However, its performance is sensitive to the amount of training data.
  • Handcrafted Features Offer Robustness: MFCCs-T/NT, while slightly less performant than CNN-T/NT, showed less degradation when training data was limited. This makes them a valuable alternative when large labeled datasets are unavailable.
  • "In-the-Wild" is Significantly Harder: Performance consistently dropped when testing on "wild" data compared to laboratory data, highlighting the challenges of real-world monitoring.
  • Percentage of Tremor Time is a Promising Biomarker: The proposed metric of estimating the percentage of tremor over longer periods (minutes to hours) proved to be a reliable and clinically relevant indicator, correlating well with ground truth and patient self-reports.

Key Results

  • The CNN-T/NT + MLP system achieved the highest performance on laboratory data, with an AUC of 0.887 and an FPR of 0.30 at a 0.90 TPR.
  • This system significantly outperformed previous state-of-the-art methods on the LAB data, including those from Pulliam et al. [6], Hammerla et al. [7], and Zhang et al. [4].
  • When predicting patient self-assessments (weak labels), the proposed system showed slightly higher correlations than Zhang et al. [4], though the differences were not statistically significant.
  • The system demonstrated good accuracy in estimating the percentage of tremor time, with mean errors of 4.1% for LAB data and 9.1% for WILD data. These errors are lower than those reported in recent studies for similar biomarkers.

Limitations and Future Work

  • Small Sample Size: The study acknowledges that the small number of subjects is a limitation. While they used Leave-One-Subject-Out (LOSO) cross-validation to mitigate this, a larger, more diverse dataset would be beneficial.
  • Limited Scope: The current system focuses solely on tremor detection (presence/absence) and does not quantify tremor intensity or type (resting, postural, kinetic).
  • Label Quality in the Wild: Obtaining accurate labels in uncontrolled environments remains a significant challenge. Future work could explore personalization algorithms to improve detection from unlabeled user-specific data.
  • Supervised Tremor Therapy Monitoring: The authors suggest future research into using their methods to supervise tremor in PD patients before and after Levodopa therapy, aiming to obtain objective biomarkers for medication impact.
  • Combining LAB and WILD Data: Investigating frameworks that can simultaneously train on accurate (LAB) and weak (WILD) labels could leverage the variability of wild data while mitigating poor label quality.

Summary updated

Abstract

Continuous in-home monitoring of Parkinson’s Disease (PD) symptoms might allow improvements in assessment of disease progression and treatment effects. As a first step towards this goal, we evaluate the feasibility of a wrist-worn wearable accelerometer system to detect PD tremor in the wild (uncontrolled scenarios). We evaluate the performance of several feature sets and classification algorithms for robust PD tremor detection in laboratory and wild settings. We report results for both laboratory data with accurate labels and wild data with weak labels. The best performance was obtained using a combination of a pre-processing module to extract information from the tremor spectrum (based on non-negative factorization) and a deep neural network for learning relevant features and detecting tremor segments. We show how the proposed method is able to predict patient self-report measures, and we propose a new metric for monitoring PD tremor (i.e., percentage of tremor over long periods of time), which may be easier to estimate the start and end time points of each tremor event while still providing clinically useful information.
BibTeX
@article{segundo2020parkinson,
  author = {Rubén San Segundo and Ada Zhang and Alexander Cebulla and Stanislav Panev and Griffin Tabor and Katelyn Stebbins and Robyn E. Massa and Andrew S. Whitford and Fernando De la Torre and Jessica K. Hodgins},
  title = {{Parkinson's Disease Tremor Detection in the Wild Using Wearable Accelerometers}},
  journal = {Sensors},
  year = {2020},
  doi = {10.3390/s20205817},
}
Figure