Human activity recognition from textile electrocardiograms.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Textile sensors offer unobtrusive, continuous physiological monitoring.
• Textile electrocardiography (ECG) sensors are emerging for physical activity monitoring.
• Exercise significantly impacts heart rate, making ECG valuable for activity recognition.

Purpose of the Study:

• To evaluate the effectiveness of a single-lead ECG signal from a non-medical-grade shirt for human activity recognition (HAR).
• To assess the performance of machine learning models in classifying activities based on wearable ECG data.

Main Methods:

• Collected 228 hours of ECG data from 10 volunteers during four activities (sleeping, sitting, walking, running).
• Processed ECG signals using Short-time Fourier Transform (STFT) to create spectrograms.
• Utilized a convolutional neural network (CNN) for activity classification using spectrogram images.

Main Results:

• The best CNN configuration achieved 77% accuracy and a 73% F1-Score on the test set.
• Leave-one-subject-out cross-validation yielded F1-Scores between 41% and 80%, indicating subject-specific performance variations.
• The study demonstrates that single-lead, wearable ECG data holds informative value for HAR.

Conclusions:

• Wearable ECG shirts provide valuable data for human activity recognition.
• Further research will explore multi-sensor fusion from smart shirts to enhance HAR accuracy.
• Textile ECG sensors represent a promising technology for continuous health and activity monitoring.