Predicting Deterioration from Wearable Sensor Data in People with Mild COVID-19

Area of Science:

• Biomedical Engineering
• Infectious Disease Modeling
• Digital Health

Background:

• Coronavirus disease (COVID-19) continues to spread globally, causing significant mortality.
• Mild COVID-19 cases can unexpectedly progress to severe illness, necessitating proactive monitoring during isolation.

Purpose of the Study:

• To develop and validate a machine learning-based model for predicting deterioration in mild COVID-19 patients during their isolation period.
• To leverage data from wearable devices and clinical questionnaires for early warning.

Main Methods:

• Collected vital signs from wearable sensors and clinical questionnaires from COVID-19 patients.
• Developed machine learning models using a derivation cohort (n=50) and validated with an external cohort (n=181).
• Evaluated model performance based on prediction lead time and area under the receiver characteristic curve (AUC).

Main Results:

• The deterioration prediction model achieved an AUC of 0.99 for 10-minute advance prediction.
• The model demonstrated an AUC of 0.84 for 8-hour advance prediction.
• Key predictive variables varied based on the prediction time horizon.

Conclusions:

• Machine learning models utilizing wearable sensor data and self-reported symptoms can efficiently monitor COVID-19 patients for deterioration.
• This approach enables timely intervention for patients with mild COVID-19, potentially improving outcomes.
• Wearable technology offers a scalable solution for remote patient monitoring in infectious disease outbreaks.