Enhancing COVID-19 Screening Models With Epidemiological and Mobility Features: Machine-Learning Model Study

Area of Science:

• Epidemiology
• Machine Learning
• Public Health

Background:

• Post-COVID-19 pandemic research surged for patient screening using symptom data and machine learning (ML).
• Crucial data on patient trajectories and epidemiological conditions remained underutilized in these ML models.
• Existing ML models for COVID-19 screening often lacked comprehensive data integration.

Purpose of the Study:

• To enhance ML model performance for COVID-19 screening.
• To integrate patient symptom data with mobility and epidemic information.
• To improve the accuracy of infectious disease diagnosis through data enrichment.

Main Methods:

• Collected daily self-reported symptoms, location, and test results from 48,798 individuals via a smartphone app.
• Combined app data with Our World in Data and national epidemic information.
• Trained five ML models (logistic regression, XGBoost, LightGBM, TabNet, Google AutoML) to classify COVID-19 infection status.

Main Results:

• Integrating mobility and epidemic data significantly improved all five ML models' performance.
• The area under the receiver operating characteristic curve (AUC) increased from 0.8712 to 0.9104 with the addition of external data.
• External data sources demonstrably enhance the performance of ML models for disease screening.

Conclusions:

• Mobility and epidemic data, combined with symptom data, can significantly improve ML model accuracy for COVID-19 diagnosis.
• Incorporating contextual information enhances the capability of screening for infectious diseases like COVID-19.
• This approach offers a more robust method for public health surveillance and patient screening.