COVID-19 mortality risk assessment: An international multi-center study.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Epidemiology

Background:

• Accurate prediction of mortality risk in hospitalized COVID-19 patients is crucial for effective patient management and resource allocation.
• Existing risk stratification tools may not fully capture the complexity of COVID-19 outcomes.

Purpose of the Study:

• To develop and validate a data-driven, personalized mortality risk calculator for hospitalized COVID-19 patients.
• To leverage machine learning for accurate mortality prediction using readily available clinical data.

Main Methods:

• Utilized de-identified data from 3,927 COVID-19 positive patients across six centers and 33 hospitals.
• Developed the COVID-19 Mortality Risk (CMR) tool using the XGBoost algorithm on a derivation cohort of 3,062 patients.
• Validated the model's discrimination performance on three independent cohorts, evaluating Area Under the Curve (AUC).

Main Results:

• Identified key risk factors: increased age, low oxygen saturation (≤ 93%), elevated C-reactive protein (≥ 130 mg/L), blood urea nitrogen (≥ 18 mg/dL), and creatinine (≥ 1.2 mg/dL).
• Achieved strong predictive performance with out-of-sample AUCs of 0.90 in the derivation cohort.
• Demonstrated robust validation with AUCs of 0.92, 0.87, and 0.81 in independent European and US patient cohorts.

Conclusions:

• The CMR tool accurately predicts mortality in hospitalized COVID-19 patients using common clinical features.
• This machine learning-based risk score is the first to be trained and validated on a combined European and US cohort.
• The CMR tool is available as an online application and is currently in clinical use, aiding in patient management.