AccNet24: A deep learning framework for classifying 24-hour activity behaviours from wrist-worn accelerometer data...

Area of Science:

• Wearable technology and sensor data analysis.
• Machine learning and deep learning applications in health.
• Biomedical signal processing and human activity recognition.

Background:

• Accurate classification of 24-hour activity behaviors from wearable accelerometry data remains a significant challenge.
• Existing machine learning methods often struggle with the complexity of raw accelerometry signals.
• Developing robust and precise activity recognition systems is crucial for health monitoring and behavioral analysis.

Purpose of the Study:

• To develop and validate a novel deep learning framework for classifying 24-hour activity behaviors using wrist-worn accelerometers.
• To compare the performance of the deep learning framework against traditional machine learning algorithms.
• To establish a new standard for accurate activity prediction from accelerometry data.

Main Methods:

• A deep learning framework, AccNet24, was developed using an openly available dataset of wrist-based accelerometry data.
• Raw acceleration signals were converted into images, and deep features were extracted using transfer learning.
• A bidirectional long short-term memory (BiLSTM) network classified activities into sleep, sedentary behavior, light-intensity physical activity (LPA), and moderate-to-vigorous physical activity (MVPA).
• Performance was compared against five traditional machine learning classifiers using hand-crafted features.

Main Results:

• AccNet24 achieved consistently high accuracy (>95%) in classifying all four activity behavior categories.
• The deep learning framework significantly outperformed traditional machine learning algorithms by 16%-30% on unseen data.
• AccNet24 demonstrated superior performance in distinguishing between sleep, sedentary behavior, LPA, and MVPA.

Conclusions:

• Deep learning techniques, particularly AccNet24 utilizing signal-to-image conversion and BiLSTM, offer a highly accurate solution for 24-hour activity behavior classification.
• The findings suggest that deep learning is a promising direction for the next generation of accelerometry-based activity prediction.
• Accurate classification of daily activities from wearable sensors can enhance health monitoring and personalized interventions.