Wearable IMU-Based Human Activity Recognition Algorithm for Clinical Balance Assessment Using 1D-CNN and GRU Ensemble Model

Yeon-Wook Kim, Joa Kyunglim, Han-young Jeong, Sangmin LeeSensors 2021 · DOI

TL;DR

This paper presents a novel deep learning approach for automatically scoring the Berg Balance Scale (BBS) using wearable Inertial Measurement Units (IMUs). By combining 1D-CNN and GRU networks in a stacking ensemble, they achieve a remarkable 98.4% accuracy, significantly improving upon previous machine learning methods and offering a more efficient and objective clinical balance assessment tool.

What Problem Does This Paper Address?

Assessing balance is crucial for elderly, brain-damaged, and rehabilitation patients to prevent falls and further injuries. The Berg Balance Scale (BBS) is a widely used clinical test for this purpose, involving 14 functional tasks. However, manual scoring by therapists can be subjective and time-consuming. Existing automated scoring methods often rely on traditional machine learning, requiring extensive manual feature engineering, which is labor-intensive and prone to human error. This paper aims to bridge this gap by developing a deep learning-based auto-scoring algorithm that is more accurate, efficient, and less reliant on manual feature extraction.

Methods

The core of this work lies in a deep learning ensemble model designed for multivariate time-series data from IMUs. The proposed method involves several key steps:

  • Preprocessing: Raw IMU data undergoes noise depression, normalization, and zero-padding.
  • Data Down-Sampling: To improve computational efficiency and reduce power consumption for wearable devices, the authors investigated down-sampling the IMU data. They analyzed the frequency components of the data and determined that a sampling rate of 20 Hz was optimal, preserving most of the information while significantly reducing data volume.
  • Data Augmentation: To address the inherent data imbalance in clinical datasets (where some scores have fewer samples than others), an over-sampling technique was employed. This process generates synthetic data points to equalize the number of samples across different BBS task scores, preventing model bias towards majority classes.
  • Deep Learning Ensemble Model: The authors explored various combinations of 1D-CNN and GRU architectures. The optimal model identified was a stacking ensemble consisting of two 1D-CNN "heads" and one GRU "head." These heads extract features from the IMU data, which are then concatenated and fed into a simple dense layer acting as a "meta-learner" for the final classification. This ensemble approach leverages the strengths of both CNNs (for spatial feature extraction) and GRUs (for temporal sequence modeling).

Datasets

The study utilized data collected from IMU sensors worn by:

  • 53 patients with various brain diseases (cerebral infarction, cerebral hemorrhage, brain atrophy, or brain embolism) who were at risk of falling.
  • 3 healthy participants who performed all BBS tasks with scores from 0 to 4 to supplement the patient data and address data imbalance.

The total experimental data recorded was equivalent to 78 experimental datasets. This dataset was used for both training and evaluating the proposed model. The authors note that the dataset from a previous study [3, 27] was used as a basis for comparison.

Key Insights and Ideas

  • Deep Learning for Objective Scoring: The paper demonstrates that deep learning, specifically a 1D-CNN and GRU ensemble, can effectively automate the scoring of the BBS, moving beyond subjective human assessment.
  • The Power of Ensemble Learning: Combining multiple specialized models (1D-CNNs and GRU) in a stacking ensemble proved more effective than individual models, allowing for a more robust and comprehensive feature extraction.
  • Data Preprocessing is Crucial: Both down-sampling and data augmentation played significant roles in improving model efficiency and accuracy. Down-sampling reduced computational load without sacrificing performance, while augmentation tackled data imbalance, leading to better generalization.
  • Optimal Sampling Rate for Wearables: Identifying a specific down-sampling rate (20 Hz) that balances information content, computational efficiency, and power consumption is a practical insight for wearable health monitoring devices.
  • Data Imbalance is a Major Hurdle: The study highlights how data imbalance in clinical datasets can severely impact model performance and emphasizes the necessity of techniques like over-sampling to mitigate this.

Key Results

  • High Accuracy: The developed stacking ensemble model achieved a maximum average accuracy of 98.4% across all 14 BBS tasks.
  • Improved Efficiency: Down-sampling the IMU data from 100 Hz to 20 Hz reduced the data volume by 80% and training time by 67.7% with only a minor decrease in accuracy (0.2%).
  • Superior Performance: The proposed model's average accuracy (98.4%) significantly outperformed previous studies, including Badura's MLP (80.6%) and Kim's SVM (93.2%) by approximately 18% and 5%, respectively.
  • Enhanced Stability: The standard deviation of accuracy for the proposed model was a mere 0.7%, indicating high consistency across all BBS tasks, compared to 10.9% and 9.1% for the previous models.

Limitations and Future Work

  • Computational Cost of Deep Learning: While deep learning automates feature extraction, it can be computationally intensive. The authors acknowledge this and plan to explore attention mechanisms to identify and exclude low-contribution sensor data, further improving computational efficiency.
  • Sensor Reduction: The current model uses data from eight IMUs. Future work aims to reduce the number of sensors required by introducing advanced deep learning techniques and signal processing algorithms.
  • Generalizability: While the model shows excellent performance on the BBS, further validation on diverse patient populations and other balance assessment tasks would be beneficial.
  • Real-time Implementation: The paper focuses on offline analysis. Adapting the model for real-time scoring on wearable devices would be a valuable next step.

Summary updated

Abstract

In this study, a wearable inertial measurement unit system was introduced to assess patients via the Berg balance scale (BBS), a clinical test for balance assessment. For this purpose, an automatic scoring algorithm was developed. The principal aim of this study is to improve the performance of the machine-learning-based method by introducing a deep-learning algorithm. A one-dimensional (1D) convolutional neural network (CNN) and a gated recurrent unit (GRU) that shows good performance in multivariate time-series data were used as model components to find the optimal ensemble model. Various structures were tested, and a stacking ensemble model with a simple meta-learner after two 1D-CNN heads and one GRU head showed the best performance. Additionally, model performance was enhanced by improving the dataset via preprocessing. The data were down sampled, an appropriate sampling rate was found, and the training and evaluation times of the model were improved. Using an augmentation process, the data imbalance problem was solved, and model accuracy was improved. The maximum accuracy of 14 BBS tasks using the model was 98.4%, which is superior to the results of previous studies.
BibTeX
@inproceedings{kim2021wearable,
  author = {Yeon-Wook Kim and Joa Kyunglim and Han-young Jeong and Sangmin Lee},
  title = {{Wearable IMU-Based Human Activity Recognition Algorithm for Clinical Balance Assessment Using 1D-CNN and GRU Ensemble Model}},
  booktitle = {Sensors},
  year = {2021},
  doi = {10.3390/s21227628},
}
Figure