Passive detection of COVID-19 with wearable sensors and explainable machine learning algorithms

Area of Science:

• Digital health
• Infectious disease surveillance
• Machine learning in medicine

Background:

• Wearable sensor data, including heart rate variability and peripheral temperature, shows potential for identifying COVID-19 infection.
• Previous studies indicate correlations between sensor data and COVID-19 symptoms or severity.
• The need for scalable, objective methods for COVID-19 detection is critical, especially when self-reported symptoms are absent.

Purpose of the Study:

• To develop and validate an explainable gradient boosting prediction model for COVID-19 detection using smartwatch sensor data.
• To assess the model's performance in identifying infection, with and without self-reported symptoms.
• To evaluate the model's ability to adapt to asymptomatic cases and passively collected data.

Main Methods:

• Development of an explainable gradient boosting model based on decision trees.
• Utilized data from 38,911 individuals, including nasopharyngeal PCR swab test results for COVID-19.
• Analysis included symptomatic and asymptomatic individuals, with and without self-reported symptoms, and data collected before the test date.

Main Results:

• The model achieved an Area Under the Curve (AUC) of 0.83 in symptomatic individuals and 0.78 when considering only pre-test data.
• For all individuals, excluding self-reported symptoms, the model yielded an AUC of 0.78 (0.70 for pre-test data).
• The model outperformed state-of-the-art algorithms under tested conditions.

Conclusions:

• Explainable machine learning models using passively monitored wearable data can effectively detect COVID-19 infection.
• The developed model demonstrates adaptability to varying symptom reporting and data availability.
• This approach offers a scalable platform for infectious disease detection in diverse settings.