US-Derived ICU Mortality Model Holds Up in China, but Only After Recalibration

A mortality prediction model built for ICU patients with bloodstream infection can travel from a US database to a Chinese hospital, but only after a statistical tune-up, according to a new study published in BMC Medical Informatics and Decision Making. The research, led by Yinting Zhu, Wen Song and colleagues with senior author Qiongfang Zhu, tested whether a deliberately simple five-variable model developed on American data could still identify high-risk patients in a Chinese tertiary teaching hospital.

The team built their 30-day mortality model using MIMIC-IV, a widely used intensive care database containing 1,501 patients with bloodstream infection. The result was a parsimonious model relying on five predictors that are routinely available in almost any ICU: age, the 24-hour Sequential Organ Failure Assessment (SOFA) score, the Charlson Comorbidity Index, serum potassium, and whether the bloodstream infection was fungal.

That simplicity is the point. The catch, as the new study shows, is that a model’s performance in the population where it was built says little about how it will behave elsewhere, because healthcare systems, case mix and clinical practice differ across countries.

In internal validation on the MIMIC-IV data, the simplified five-variable model achieved an area under the receiver operating characteristic curve (AUC) of 0.776, only slightly below the 0.796 achieved by the full model.

The real test came in the external cohort: 285 ICU patients with bloodstream infection admitted to a tertiary teaching hospital in China. Discrimination held up reasonably well, with an AUC of 0.764, meaning the model still ranked patients by risk about as sharply as it had in the original US population. But calibration, the agreement between predicted probabilities and observed outcomes, was poor. The model systematically underestimated 30-day mortality in the Chinese cohort, with calibration-in-the-large greater than zero, meaning predicted risks were too low on average.

To fix this, the researchers applied local logistic recalibration, a low-resource statistical adjustment that refits the intercept and slope of the model to the new population without rebuilding it from scratch. The procedure brought the calibration-in-the-large and the calibration slope closer to their target values and reduced the Brier score, a combined measure of discrimination and calibration, to 0.182. In other words, the imported model needed only a modest statistical adjustment, not a full redevelopment, to become usable in the Chinese setting.

The authors are careful about the limits of that result. Because the recalibration was estimated and evaluated in the same Chinese cohort, the reported improvement may be optimistic, and the recalibrated model requires validation in independent healthcare systems before its performance can be considered established. The external cohort was also modest in size, at 285 patients, which constrains the precision of the performance estimates.

Even so, the study offers a practical message for hospitals considering imported prediction models. A parsimonious model developed on US data retained acceptable discrimination in this Chinese ICU cohort, but it showed important calibration drift, and local logistic recalibration offered a feasible way to correct it.

Subject of Research: Cross-national transportability and local recalibration of a 30-day mortality prediction model for ICU patients with bloodstream infection

Article Title: Cross-national transportability and local adaptation of a parsimonious 30-day mortality prediction model for ICU patients with bloodstream infection: from MIMIC-IV to a Chinese tertiary hospital

Article References: Zhu, Y., Song, W., Ren, Y., Wu, Y., Chen, X., Xu, J., & Zhu, Q. (2026). Cross-national transportability and local adaptation of a parsimonious 30-day mortality prediction model for ICU patients with bloodstream infection: from MIMIC-IV to a Chinese tertiary hospital. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03929-9

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03929-9

Keywords: ICU, bloodstream infection, mortality prediction, MIMIC-IV, external validation, calibration, recalibration, SOFA score, machine learning, China, prediction model transportability, Cross-national