Transfer learning for COVID-19 predictive modeling: A multicenter study of 12 hospitals

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• COVID-19 Research

Background:

• Predicting Intensive Care Unit (ICU) admission is crucial for managing COVID-19 patients.
• Multicenter studies face challenges due to variations in patient populations and data.

Purpose of the Study:

• To apply transfer learning for predicting ICU admission in COVID-19 patients across multiple hospitals.
• To evaluate the effectiveness of transfer learning in improving model generalization and performance.

Main Methods:

• XGBoost algorithms were trained using demographic and laboratory data from 12 hospitals.
• A best-performing hospital model was identified and externally validated.
• Transfer learning was implemented by fine-tuning the best model with data from other hospitals.

Main Results:

• Significant variation in local predictive performance (AUC 0.6239-0.9410) was observed.
• External validation showed AUC below 0.7 in 6 out of 11 hospitals.
• Transfer learning, after incorporating 20 new trees, improved AUC in 9 hospitals compared to external validation.

Conclusions:

• Transfer learning effectively utilizes knowledge from high-performing hospitals.
• The method accelerates model training and enhances adaptability in hospitals with limited data or lower performance.
• This approach improves the efficiency and generalizability of predictive models in a multicenter setting.