Source-linked AI summary
Data Augmentation of Wearable Sensor Data for Parkinson's Disease Monitoring using Convolutional Neural Networks
Terry Taewoong Um, Franz Michael Josef Pfister, Daniel Pichler, Satoshi Endo, Muriel Lang, Sandra Hirche, Urban Fietzek, Dana Kulić
TL;DR
CNNs need large labeled datasets, yet PD motor-state classification from wearable sensors is difficult because data are limited and labels are noisy. The paper proposes and evaluates wearable-sensor augmentations with CNNs, finding that appropriate combinations improve accuracy from 77.52% to 86.88%.
Problem
PD motor-state classification is challenging because wearable-sensor data have limited availability, noisy labels, and large variability.
Method
The paper evaluates multiple wearable-sensor data augmentation methods with a 7-layer CNN for automatic PD motor-state monitoring.
Results
86.88% accuracy was achieved by combining rotational, permutational, and time-warping augmentation, improving the 77.52% CNN baseline.
Takeaways & Limitations
Rotation addresses sensor-pose variability, while permutation and time-warping address temporal event-location variability in windows.
Takeaways & Limitations
Observed signals can disagree with expert labels because wrist motion, fixed-window segmentation, voluntary movements, and tremor may not represent the labeled motor state.
Abstract
from arXiv · showhide
While convolutional neural networks (CNNs) have been successfully applied to many challenging classification applications, they typically require large datasets for training. When the availability of labeled data is limited, data augmentation is a critical preprocessing step for CNNs. However, data augmentation for wearable sensor data has not been deeply investigated yet. In this paper, various data augmentation methods for wearable sensor data are proposed. The proposed methods and CNNs are applied to the classification of the motor state of Parkinson's Disease patients, which is challenging due to small dataset size, noisy labels, and large intra-class variability. Appropriate augmentation improves the classification performance from 77.54\% to 86.88\%.
1 INTRODUCTION
The paper addresses the difficulty of applying CNNs to small labeled medical datasets by proposing wearable-sensor data augmentation for PD motor-state classification. It presents augmentation methods and compares them experimentally.
- CNNs perform well with large labeled datasets but are difficult to apply when medical datasets are small.Collecting and labeling large amounts of medical data is often difficult.
- Wearable-sensor augmentation must create new samples while preserving labels, which is difficult when motion intensity distinguishes classes.Scaling acceleration data may change labels because some classes are differentiated by motion intensity.
- PD motor-state classification with CNNs is challenging because of noisy labels, irrelevant motion interference, patient variability, and limited labeled data.
- The paper contributes CNN-based PD motor-state classification, wearable-sensor augmentation methods, and an experimental comparison of those methods.The study uses clinician-labeled data from 30 patients, exploiting data from 25 patients in daily-living conditions.
2 RELATED WORK
Prior work used wearable sensors and deep learning for automated PD evaluation, but wearable-sensor augmentation had not been systematically investigated. The paper studies augmentation for CNN-based PD motor-state classification.
- PD motor fluctuations include bradykinesia and dyskinesia, and accurate phenomenology evaluation helps determine medication dosage.Dopaminergic treatment may alleviate bradykinesia, while over-treatment can cause dyskinesia.
- Existing PD motor-state evaluation relies on patient self-reports and clinician visual observation.
- Wearable-sensor approaches have mostly focused on standardized clinical motor tasks, while daily-living evaluation requires substantial data.Deep learning is presented as a promising approach for handling PD data variability, but collecting large datasets is difficult.
- Wearable-sensor data augmentation had not been systematically investigated, motivating evaluation of multiple augmentation methods with CNNs.Earlier time-series work included window slicing and window warping, but the paper proposes and evaluates a broader set of methods.
3 PD MOTOR STATE CLASSIFICATION
PD motor-state sensor data exhibit label noise, intra-class variability, and overlap because recorded wrist signals can differ from expert labels. The paper proposes transformations targeting sensor placement, event timing, noise, magnitude, and event-location dependence.
- 3.1 Challenges in PD Data: Wearable signals can disagree with expert labels when the monitored hand does not reflect whole-body symptoms or fixed windows mix motor states.Voluntary movements, voluntary rest, and tremor can also make bradykinesia and dyskinesia signals appear similar.
- 3.1 Challenges in PD Data: Noisy labels and large intra-class variability create substantial overlap between motor-state classes, making classification harder with limited data.
- 3.2 Data Augmentation Methods for Wearable Sensor Data: Sensor-placement variability motivates arbitrary rotations that simulate different sensor orientations without changing labels.
- 3.2 Data Augmentation Methods for Wearable Sensor Data: Permutation divides each window into 1–5 equal segments and randomly reorders them, while time-warping smoothly changes sample timing.
- 3.2 Data Augmentation Methods for Wearable Sensor Data: Jittering simulates additive sensor noise, while magnitude-warping smoothly varies each sample’s magnitude around one.Scaling multiplies a window by a random scalar, but its description is incomplete in the supplied passage.
- 3.2 Data Augmentation Methods for Wearable Sensor Data: Cropping reduces dependence on event locations but may remove critical or event-free regions and thereby change the label.The paper applies jittering, scaling, cropping, rotating, permutating, magnitude-warping, and time-warping before CNN evaluation.
4 EXPERIMENTS
Experiments evaluate CNN-based Parkinson’s motor-state classification on wearable-sensor data from daily living, comparing augmentation strategies under limited and noisy data conditions. Rotation, permutation, and time-warping improve performance, with their combination achieving the strongest result.
- Data preparation: The dataset contains 3530 minutes of bradykinesia and dyskinesia data from 25 patients, collected in daily-living conditions and evaluated with five-fold cross-validation.Data were collected using Microsoft Band 2 and labeled in one-minute intervals by a clinical expert.
- Model: The model is a deep, sparse 7-layer CNN using strided convolutions, inter-axis accelerometer features, and global average pooling to reduce parameters.The CNN reduces inputs from 6960*3 to 48*1 and applies GAP instead of a fully connected layer.
- Augmentation results: Jittering, cropping, scaling, and magnitude-warping fail to improve classification because they introduce misleading fluctuations, discard critical information, or alter label-relevant signal intensity.These effects are especially problematic when the dataset is small or labels depend on movement intensity.
- Augmentation results: Rotation, permutation, and time-warping improve accuracy by 3.6-5.1% by addressing sensor-placement and event-location variability.Rotation is strongest among individual augmentation methods, while permutation and time-warping perturb temporal sample locations.
- Augmentation results: Rot+Perm and Rot+TimeW improve over the CNN baseline by 7.5-9.2%, while Rot+Perm+TimeW achieves 86.88% accuracy.The combined method trains more slowly but shows better generalization, consistent with a regularization effect.
5 CONCLUSION
The paper develops CNN-based automatic monitoring of Parkinson’s motor states from wearable-sensor data and proposes augmentation methods for this limited-data setting. Combining rotational and permutational augmentation improves baseline accuracy from 77.52% to 86.88%.
- Conclusion: The study develops an automatic PD motor-state classification algorithm using wearable sensor data and a 7-layer CNN.The task is challenging because of large inter-class variability, noisy labels, irrelevant motion interference, and limited data.
- Conclusion: The paper proposes and evaluates multiple data augmentation approaches for wearable-sensor datasets used in CNN-based classification.The contributions include applying CNNs to daily-living PD data and systematically comparing augmentation methods.