Objectives This study is part of the Dutch Pediatric Intensive Care Evaluation (PICE) and the aim is to determine which general mortality prediction model, PIM or PRISM, is most appropriate for use in the pediatric intensive care setting in the Netherlands.
It is a national, multi-center, prospective, observational study involving the performance of original and updated PIM and PRISM.
Methods We analyzed admissions between March 2006 and August 2007 to all PICUs in the Netherlands as part of the PICE registry.
The models were examined on their accuracy by comparing them on overall performance using the standardized mortality ratio (SMR), discrimination by calculating the area under the ROC curve (AUC) and calibration by comparing observed-predicted ratio of deaths among deciles of risk with the Hosmer-Lemeshow goodness of fit test (H-L test).
Results A total of 6755 out of 6770 admissions were included for final analysis.
Only the original PRISM showed significant overprediction of mortality with an overall SMR significantly below 1; the other models included 1 in the SMR.
The discrimination of the models was good, with best discriminating powers (AUC of 0.9) for PRISM models.
The models were poorly calibrated for the Dutch setting, H-L test resulted in p values <0.05.
Conclusion All models needed recalibration to the Dutch setting. After simple recalibration overprediction of mortality decreased and discrimination stayed the same. The PRISM is favoured because of better discriminating powers and acceptable calibration and SMR after recalibration.