Fitbeat: COVID-19 estimation based on wristband heart rate using a contrastive convolutional auto-encoder

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Epidemiology

Background:

• Remote health monitoring is crucial for managing chronic conditions like multiple sclerosis (MS).
• Detecting infectious diseases such as COVID-19 early is vital for public health and individual care.
• Wearable devices offer continuous physiological data collection for disease surveillance.

Purpose of the Study:

• To develop and evaluate a deep learning model for identifying suspected COVID-19 infection using heart rate data.
• To assess the efficacy of a contrastive convolutional auto-encoder (contrastive CAE) in detecting COVID-19 symptoms in individuals with MS.
• To compare the performance of the contrastive CAE against conventional machine learning models.

Main Methods:

• Utilized heart rate data collected remotely via Fitbit wristbands from participants in the RADAR-CNS mHealth project.
• Developed a contrastive convolutional auto-encoder (contrastive CAE) architecture integrating auto-encoder principles with contrastive loss.
• Compared the contrastive CAE with a standard convolutional neural network (CNN), a long short-term memory (LSTM) model, and a non-contrastive convolutional auto-encoder (CAE).

Main Results:

• The contrastive CAE significantly outperformed CNN, LSTM, and CAE models in detecting COVID-19 symptoms.
• Achieved an unweighted average recall of 95.3%, 100% sensitivity, and 90.6% specificity on a test set.
• Demonstrated a high area under the receiver operating characteristic curve (AUC-ROC) of 0.944, indicating robust detection with a low false alarm rate.

Conclusions:

• The contrastive CAE is a promising tool for non-invasive, remote detection of COVID-19 symptoms using heart rate data.
• This approach holds potential for early identification of infections in individuals with MS and other populations.
• Integrating advanced AI with mHealth data can enhance infectious disease surveillance and management.