Predicting COVID-19 mortality with electronic medical records

Area of Science:

• Computational epidemiology
• Clinical informatics
• Public health

Background:

• Predicting mortality after COVID-19 infection is crucial for resource allocation.
• Existing prognostic models often require extensive clinical data beyond routine electronic health records (EHRs).
• Understanding age-specific risk factors can refine predictive accuracy.

Purpose of the Study:

• To develop a predictive model for COVID-19 mortality using only pre-existing EHR data.
• To identify and analyze age-stratified risk factors for death post-COVID-19 infection.
• To assess the performance of EHR-based models compared to symptom- and lab-based models.

Main Methods:

• Utilized computational methods and clinical expertise to identify 46 potential risk factors from EHRs.
• Trained age-stratified generalized linear models (GLMs) with component-wise gradient boosting.
• Validated model performance against existing prognostic approaches.

Main Results:

• Models using only past EHR data achieved performance comparable to models requiring extensive clinical information.
• Age was the most significant predictor of COVID-19 mortality.
• A history of pneumonia emerged as a critical risk factor across all age groups.
• Specific comorbidities like diabetes with complications and certain cancers (breast, prostate) were significant for middle-aged adults (45-65).
• Pulmonary diseases (interstitial lung disease, COPD, lung cancer) and smoking history were key predictors for older adults (65-85).

Conclusions:

• Predicting COVID-19 mortality is feasible using readily available EHR data, simplifying risk assessment.
• Individualized risk scores derived from EHRs can aid in resource allocation, such as vaccine prioritization.
• Identifying age-specific risk factors enhances the precision of mortality prediction.