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

Abstract

Purpose

To use transfer learning to predict Intensive Care Unit admission (ICU) for COVID-19 patients in a multicenter setting.

Methods

We trained XGBoost algorithms using demographic and laboratory predictors recorded during hospital admissions across 12 hospitals. Our goal was to identify the best-performing hospital and evaluate its generalization capacity through external validation in the other 11 hospitals. Additionally, this best-performing hospital served as the source domain for transfer learning and was fine-tuned by incorporating new decision trees trained with local data from each of the remaining hospitals.

Results

We observed considerable variation in ICU admission outcomes across the 12 hospitals, with local predictive performance (AUC – area under the receiver operating characteristic curve) ranging from 0.6239 to 0.9410. The best-performing model's generalization capacity, evaluated through external validation, resulted in an AUC below 0.7 in 6 hospitals. By applying transfer learning and fine-tuning this best-performing model, we found that its performance stabilized after incorporating 20 new trees trained with data from the target hospitals. This adjustment improved AUC in 9 out of the 11 hospitals, compared to the results from the external validation.

Conclusions

The study showed that transfer learning effectively leverages knowledge from the best-performing hospital to initialize and accelerate model training in other hospitals with lower performance or limited sample sizes, thereby enhancing overall adaptability and efficiency.