Exploring the Possibility of Photoplethysmography-Based Human Activity Recognition Using Convolutional Neural...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Human Activity Recognition (HAR) research utilizes diverse sensors, with wearable internal sensors showing promise due to their simplicity.
• Existing HAR methods using wearable biometric signals like ECG and PPG often rely on limited datasets, hindering practical application.
• A need exists for extensive analysis with cross-subject validation to advance wearable-based HAR.

Purpose of the Study:

• To investigate the feasibility of classifying common daily activities using photoplethysmography (PPG) signals.
• To develop and evaluate a deep learning model for activity recognition based on PPG data.
• To determine an optimal signal processing window size for PPG-based HAR.

Main Methods:

• Collected PPG signals from 40 participants performing five distinct daily activities.
• Employed a deep learning architecture for activity classification.
• Conducted cross-subject cross-validation to assess model performance using accuracy, precision, recall, and F1-score.

Main Results:

• The deep learning model achieved an average test accuracy of 95.14% in distinguishing the five activities.
• Performance metrics (accuracy, precision, recall, F1-score) confirmed the model's effectiveness.
• An optimal window size for signal processing was identified through comprehensive evaluation.

Conclusions:

• PPG signals are highly effective for practical human activity recognition using deep learning.
• This technique offers potential for integrated behavioral and health monitoring in healthcare and fitness applications.
• A single biometric signal can concurrently analyze activity and health data, paving the way for novel applications.