AI-derived CT biomarker score for robust COVID-19 mortality prediction across multiple waves and regions using...

Area of Science:

• Medical Informatics
• Radiology
• Infectious Diseases

Background:

• Complex models often limit clinical application for predicting COVID-19 mortality.
• There is a need for simple, interpretable predictive tools using routinely available data.

Purpose of the Study:

• To develop a simple, interpretable model for predicting COVID-19 mortality at admission.
• To provide a statistically robust framework for clinical use, managing model uncertainty.

Main Methods:

• Developed and validated the M3-score model using machine learning on COVID-19 patient data.
• Incorporated age, white blood cell (WBC) count, and AI-derived total lung involvement (TOTAL_AI) from CT scans.
• Compared diagnostic performance and categorized probabilities into likelihood ratio (LR) intervals.

Main Results:

• The M3-score achieved strong performance in training (AUC 0.903) and useful performance in external validation (AUC 0.826).
• Categorized probabilities into actionable LR intervals (e.g., unlikely LR 0.13, likely LR 8.19).
• Demonstrated potential generalizability and temporal/geographical robustness, with some expected real-world variability.

Conclusions:

• The parsimonious M3-score is an interpretable tool for predicting in-hospital COVID-19 mortality.
• AI-based CT quantification integrated with clinical data enhances predictive capability.
• Further extensive international validation is needed before widespread clinical adoption.