Automatic physical activity recognition using multichannel, fusion CNN-BiGRU-Bahdanauattention networks

Area of Science:

• Ubiquitous computing
• Biomedical engineering
• Machine learning for healthcare

Background:

• Human activity recognition (HAR) is vital for precision medicine in chronic disease management, especially for diabetes.
• Current deep learning models face challenges in feature extraction and action segmentation with time-series data.
• Accurate physical activity tracking is crucial for effective diabetes therapy and blood glucose control.

Purpose of the Study:

• To develop an advanced multichannel fusion model for enhanced human activity recognition.
• To improve the accuracy and efficiency of classifying physical activities in time-series data for diabetes management.
• To evaluate the performance of various machine learning classifiers for HAR tasks.

Main Methods:

• A multichannel fusion model integrating a convolutional neural network (CNN) and a bidirectional gated recurrent unit (Bi-GRU) with a Bahdanau attention mechanism.
• Utilized extra trees classifier for final classification after feature extraction and temporal learning.
• Model performance evaluated on the UCI-HAR dataset with cross-validation.

Main Results:

• The proposed fusion model achieved high accuracy (99.52%), precision (99.56%), recall (99.55%), and F1 score (99.55%) when combined with the extra trees classifier.
• Demonstrated superior performance compared to existing models in recognizing various physical activity types.
• Successfully addressed challenges in feature extraction and temporal relationship learning.

Conclusions:

• The developed multichannel fusion model offers a significant advancement in human activity recognition for diabetes management.
• The integration of CNN, Bi-GRU, attention mechanism, and extra trees classifier provides a robust solution for precise physical activity classification.
• This approach holds promise for enhancing precision medicine strategies in managing chronic illnesses like diabetes.