We studied the performance of PIM-2 in a large sample of patients admitted during 2009–2011 to three PICUs and one CICU of a large tertiary pediatric hospital in Italy. The high rates of completion achieved confirm that the PIM-2 score is easy to apply and includes variables easily available at admission to intensive care. PIM-2 appeared to have a fair discriminating ability, with an overall AUC of 0.8. However, calibration as assessed by Hosmer-Lemeshow goodness of fit test was less satisfactory, with statistically significant overprediction of deaths in the total sample, mainly due to the over two-fold overprediction in the highest risk group.
Similar findings were reported by other Authors [11, 15–17], and variably attributed to differences in the characteristics of the study population, small sample size, particularly for number of deaths, and improvement in the quality of intensive care making the original PIM-2 equation, developed about ten years ago, no longer fully appropriate.
Poor calibration by risk groups was not observed in the multicentric study validating PIM-2 for the Italian population . This validation study was conducted in 2004–2005, and we cannot exclude that quality of care has further improved since then. Additionally, that study included 18 between pediatric and general hospitals with a pediatric ICU, and Unit organization and care delivered are likely to have been less homogeneous in a such large sample of different hospitals than between the ICUs of a single hospital entirely dedicated to the care of children.
In our study, overprediction of deaths was however particularly evident among cardiac surgical admissions, where the number of expected deaths in the highest risk group was approximately four times higher than observed. These results are in agreement with those found by Czaja et al. in their study of cardiac surgical pediatric patients in 55 US PICU in 2005–07 . The mortality reported by this study (3.4%) was very close to the 3.3% rate observed in our population of children who underwent cardiac procedures. Also discrimination was similar in the two studies, with AUC values of 0.72 (CI 0.63-0.82) in Italy, and 0.80 (95% CI, 0.77– 0.83) in US, while standardised mortality rates were 0.5 (CI 0.4-0.7) and 0.8 (0.7–0.9), respectively.
In our hospital cardiac and cardiac surgical patients are treated in a separate intensive care Unit (CICU), characterized by highly specialized dedicated staff and high activity volume. Consistency between our results and those by Czaja et al.  suggests that PIM-2 calibration for cardiac surgical patients may indeed be sub-optimal. The development of a specific algorithm for predicting the risk of death in this highly selected category of patients would thus be helpful.
At the light of the statistically significant differences in observed mortality found between Units, main reasons of admission to intensive care and patient age class, we performed a multivariable analysis to simultaneously explore the effect of these variables on risk of death, adjusting for PIM2 score. In fact, SMRs did vary by Unit, and different Units admitted patients with different age distribution and main reason for admission. It was then of interest to investigate whether in our patient population these variables and the Unit of admission may have a significant influence on mortality in excess to the proportion predicted by PIM-2 score.
We found that, independently from PIM-2, main reason for admission and age remain significant predictors of the observed mortality, with lower risk in cardiac surgical patients and in children from 1 month to 12 years of age, and higher risk in medical patients and in babies < 1 month of age or adolescents. It is interesting to note that the two Units with SMRs significantly lower than 1 admitted mostly patients in the cardiac surgical group (CICU) or from 1 month to 12 years of age (PICU1), confirming the importance of differences of patient case-mix.
The lower risk in cardiac surgical patients supports the importance of cardiac focused risk adjustment. The role of age on risk of death needs to be further explored. In our study, the observed mortality was higher in the first month of life and in adolescents. In these age groups, the PIM-2 standardized mortality was very close to the observed mortality, while it was significantly lower in children from 1 month to 12 years of age. The highest risk of dying in newborns, neonates and adolescents observed in multivariate analysis could be due to differences in outcome independent from patients severity at PICU/CICU admission. In fact, both patients < 1 month of age and adolescents have specific characteristics, and could require a more targeted care. Babies < 1 month of age requiring intensive care can also be treated in neonatal Intensive Care Units (NICU). It would thus be relevant to compare PIM-2 performance and SMRs in newborns and neonates admitted in NICUs versus PICUs. Teenagers are different from both adults and pre-adolescent children and, as such, require special attention in the manner in which healthcare services are provided. However, there are no standards to guide the model of care that should be provided specifically for teenagers in intensive care .
Potential limitations of this study should be taken into account. Data regarding the main diagnosis were derived from patients’ medical record at discharge, and misclassification cannot be completely excluded. However, in our hospital all clinical records are independently reviewed at discharge by a second physician, who verifies the appropriateness of diagnoses and procedures coding. We derived main reason for PICU/CICU admission from routinely collected data, and more detailed information on reason for intensive care admission was not available. Data related to PIM-2 were collected prospectively by intensive care physicians at time of admission, and rates of completeness were very high both on total patients (91%) and deaths (95%).