PCovNet+: A CNN-VAE anomaly detection framework with LSTM embeddings for smartwatch-based COVID-19 detection

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Epidemiology

Background:

• The COVID-19 pandemic highlighted global unpreparedness for highly contagious pathogens.
• Technological advancements in healthcare are crucial for future pandemic response and monitoring.
• Wearable devices offer a potential avenue for continuous health surveillance.

Purpose of the Study:

• To propose PCovNet+, a deep learning framework for detecting COVID-19 infection using Resting Heart Rate (RHR) data from smartwatches and fitness trackers.
• To address data scarcity challenges in personalized health monitoring models through pre-training.
• To evaluate the efficacy of PCovNet+ as a secondary diagnostic tool for early infection detection.

Main Methods:

• Development of PCovNet+, a deep learning framework integrating a convolutional neural network (CNN)-based variational autoencoder (VAE) with a long short-term memory (LSTM) network.
• Utilizing latent space embeddings from the VAE-LSTM model for anomaly detection in RHR.
• Employing pre-training on data from healthy subjects to enhance model performance and generalization.

Main Results:

• PCovNet+ achieved high performance in anomalous RHR detection for COVID-19 infected subjects, with precision, recall, F-beta, and F-1 scores of 0.993, 0.534, 0.9849, and 0.6932, respectively.
• The framework successfully detected COVID-19 infection in 74% of subjects, including 47% presymptomatic and 27% post-symptomatic detection.
• Demonstrated significant improvement in detection metrics compared to existing literature.

Conclusions:

• PCovNet+ shows significant potential as a non-invasive, secondary diagnostic tool for COVID-19 detection.
• Continuous health monitoring via wearable RHR data can aid in early identification of infections.
• The framework supports enhanced pandemic preparedness through early warning systems and contact tracing capabilities.