Classification of Patient Recovery From COVID-19 Symptoms Using Consumer Wearables and Machine Learning

Area of Science:

• Digital Health
• Machine Learning in Medicine
• Wearable Technology

Background:

• Current COVID-19 remote monitoring heavily depends on manual symptom reporting, which is limited by patient compliance.
• Manual reporting can lead to incomplete or inaccurate data for assessing patient recovery.
• There is a need for automated, objective methods for remote patient monitoring.

Purpose of the Study:

• To develop and evaluate a machine learning (ML)-based remote monitoring system (eCOVID) for estimating COVID-19 patient recovery.
• To utilize automatically collected wearable device data as an alternative to manual symptom reporting.
• To assess the efficacy of ML models in predicting patient recovery status.

Main Methods:

• Deployed the eCOVID system in two COVID-19 telemedicine clinics, collecting data from Garmin wearables and a symptom tracker mobile app.
• Collected data included patient vitals, lifestyle information, and self-reported symptoms for daily recovery status labeling.
• Developed and evaluated ML-based binary classifiers, including Random Forest (RF), using leave-one-subject-out cross-validation.

Main Results:

• The Random Forest (RF) model demonstrated the highest performance among the evaluated machine learning approaches.
• A personalized RF model, utilizing weighted bootstrap aggregation, achieved an F1-score of 0.88.
• The study confirmed the feasibility of using wearable data for automated recovery status estimation.

Conclusions:

• Machine learning-assisted remote monitoring with wearable data can effectively supplement or replace manual symptom tracking in COVID-19 patients.
• Automated data collection from wearables enhances the reliability and objectivity of remote patient monitoring.
• This approach addresses the limitations of patient compliance in traditional remote monitoring methods.