Multivariable mortality risk prediction using machine learning for COVID-19 patients at admission (AICOVID)

Area of Science:

• Medical Informatics
• Public Health
• Clinical Medicine

Background:

• Early identification of Coronavirus disease 2019 (COVID-19) patients at high risk of mortality is crucial for effective resource allocation and timely interventions.
• Developing accurate predictive models can significantly improve patient triage, management strategies, and overall outcomes.

Purpose of the Study:

• To develop and validate individualized mortality risk scores for COVID-19 patients using clinical and laboratory data available at admission.
• To determine the probability of death at 7 and 28 days post-admission.

Main Methods:

• Utilized a Machine Learning (ML) approach, specifically the eXtreme Gradient Boosting (XGB) Algorithm, combined with Cox Proportional Hazard Model for 'time to event' analysis.
• Collected and analyzed anonymized clinical and laboratory data from 1393 COVID-19 patients across six hospital centers (April-July 2020).
• Prospective validation was performed on a cohort of 977 patients (July-October 2020).

Main Results:

• Identified significant predictors of mortality including Age, Male Gender, Respiratory Distress, Diabetes Mellitus, Chronic Kidney Disease, Coronary Artery Disease, elevated respiratory rate, low oxygen saturation, decreased Lymphocyte percentage, elevated INR, LDH, and Ferritin.
• The developed model achieved an AUC ROC Score of 0.8685 and an Accuracy Score of 96.89% in the initial cohort.
• The validation cohort demonstrated strong performance with an AUC of 0.782 and Accuracy of 0.93.

Conclusions:

• The developed ML model accurately predicts COVID-19 mortality risk based on initial clinical and laboratory parameters.
• This 'time to event' prediction model offers valuable insights for clinical decision-making and patient management.
• The study highlights the potential of ML in enhancing early risk stratification for COVID-19 patients.