Ensemble learning for poor prognosis predictions: A case study on SARS-CoV-2.

Area of Science:

• Epidemiology
• Machine Learning
• Biostatistics

Background:

• Risk prediction models are crucial for clinical decisions but often lack generalizability across diverse populations.
• The severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) pandemic highlighted the need for robust prediction models adaptable to varied prognoses.
• Existing models developed from single cohorts struggle with population-level consistency.

Purpose of the Study:

• To address the challenge of inconsistent performance of individual risk prediction models.
• To develop and validate an ensemble learning framework for synergizing existing COVID-19 prediction models.
• To achieve personalized risk predictions for individual patients using a combined model approach.

Main Methods:

• Selected and reimplemented 7 existing COVID-19 prediction models from diverse cohorts.
• Developed a novel ensemble learning framework to integrate these models.
• Validated 8 models (7 individual + 1 ensemble) on 4 international cohorts (UK and China; N=5394) assessing discrimination, calibration, and clinical usefulness.

Main Results:

• Individual models exhibited variable performance across different cohorts.
• The proposed ensemble model consistently achieved superior performance in discrimination, calibration, and clinical usefulness.
• All models performed better on China cohorts compared to UK cohorts, indicating country-specific performance disparities.

Conclusions:

• Ensemble learning effectively synergizes diverse prediction models to create a robust, high-performing model.
• The ensemble approach enhances predictive accuracy by selecting the most competent individual models for specific patients.
• Early collection of blood parameters and physiological measurements may improve predictive power, warranting further investigation.