CovidRhythm: A Deep Learning Model for Passive Prediction of Covid-19 Using Biobehavioral Rhythms Derived From...

Area of Science:

• Biomedical Engineering
• Data Science
• Computational Biology

Background:

• The SARS-CoV-2 virus causes physiological disruptions, including alterations in heart rate and rest-activity patterns.
• Early detection of Covid-19 is crucial for effective management and containment.
• Consumer-grade wearables offer a non-invasive method for continuous physiological monitoring.

Purpose of the Study:

• To develop and validate a deep learning model, CovidRhythm, for detecting Covid-19 using wearable sensor data.
• To investigate the utility of physiological and biobehavioral rhythm features for early Covid-19 detection.
• To assess the performance of a novel Gated Recurrent Unit (GRU) Network with Multi-Head Self-Attention (MHSA) in identifying Covid-19.

Main Methods:

• A novel deep learning model, CovidRhythm (GRU-MHSA), was developed to predict Covid-19.
• Features were extracted from heart rate and activity (steps) data collected via smart wearables.
• Biobehavioral rhythms were modeled using parameters like mesor, amplitude, and acrophase, alongside sensor-derived features.

Main Results:

• CovidRhythm achieved an AUC-ROC of 0.79, demonstrating effective discrimination between Covid-positive patients and healthy controls.
• Rhythmic features were highly predictive of Covid-19 infection, outperforming sensor features alone.
• Circadian rest-activity rhythms showed the most significant disruption in infected individuals.

Conclusions:

• Deep learning models analyzing biobehavioral rhythms from wearables can facilitate timely Covid-19 detection.
• This study presents the first use of deep learning and biobehavioral rhythms from consumer wearables for Covid-19 detection.
• The findings highlight the potential of smart wearables for non-invasive, early detection of infectious diseases.