Patient-specific COVID-19 resource utilization prediction using fusion AI model

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Epidemiology

Background:

• The COVID-19 pandemic strained healthcare resources, necessitating predictive models for resource allocation.
• Electronic Medical Records (EMR) contain valuable data for predicting patient outcomes and resource utilization.

Purpose of the Study:

• To assess the feasibility of predicting COVID-19 hospitalization needs at the time of RT-PCR testing.
• To evaluate machine learning models using pre-test EMR data for hospitalization prediction.

Main Methods:

• Utilized EMR data from 3194 COVID-19 positive patients across 12 Emory Healthcare centers (Feb-Sep 2020).
• Employed five EMR modalities: demographics, medications, procedures, comorbidities, and lab results.
• Developed and compared predictive models using early, middle, and late data fusion techniques.

Main Results:

• The early fusion model achieved the highest predictive performance with an 84% F1-score (CI 82.1-86.1).
• Omitting long-term medical history reduced predictive performance by 6%.
• Key predictors included cardiovascular disease history, recent ER visits, and demographics.

Conclusions:

• Fusion modeling integrating medical history and current data effectively forecasts COVID-19 hospitalization risk at diagnosis.
• This approach supports proactive healthcare resource planning and patient management.