Development and validation of a bedside prediction score for nosocomial sepsis in the pediatric ICU: a prospective observational cohort study

Background

Diagnosis of nosocomial sepsis (NS) is a challenge in every pediatric ICU [1, 2]. There are very few studies on NS prediction in children, and those existing [3–8] have studied risk factors with emphasis on admission variables. In contrast, our study attempts to simplify the decision-making process using a dynamic scoring system based on objective criteria.

Methods

This was a prospective study where 428 consecutive admissions, aged 1 month to 12 years, between January and October 2011 with PICU stay >48 hours, were enrolled and followed-up during their ICU stay and 72 hours thereafter. Occurrence of culture-positive nosocomial infections and relevant details were recorded. Patients with and without nosocomial sepsis were compared by chi-square test or Fisher's exact test for categorical and unpaired t test or Kruskal-Wallis for continuous variables. Significant predictors of NS (P < 0.05 on univariate analysis) were included in the binary backward stepwise logistic regression. The resultant derivation model's discrimination and calibration were assessed using the receiver operator characteristic (ROC) curve and Hosmer-Lemeshow test, respectively. The final model was transformed into a score, based on the regression coefficients. For internal validation, bootstrapping and shrinkage coefficients were used.

Results

Of the 428 enrolled, 17 were excluded (malignancies (14 cases), burns (one case), polytrauma (one case) and missing data (one case)). A total of 151 episodes (23.1%; 95 out of 411 children) of culture-positive NS were seen giving an incidence rate of 4.5 per 100 patient-days. Age, PRISM III score, device utilization, albumin, immunomodulator and prior antibiotic use, and intubation were significant independent predictors on multivariate analysis (Table 1). This model had an AUC-ROC of 0.87 (Figure 1). Also, the Hosmer-Lemeshow chi-square was 5.06 (P = 0.75) indicating good fit of the model. Based on the regression coefficients, a pediatric nosocomial sepsis prediction score (Pe-NoSeP) was developed. Probability cutoffs versus sensitivity and specificity plotting showed a cutoff of 0.38 corresponding to a score of 15. The sensitivity, specificity, positive predictive value, negative predictive value and positive likelihood ratio at this cutoff were 79.1%, 79.1%, 61.6%, 89.9% and 3.76, respectively. The accuracy of the model was 79.3% and reduced classification errors from 29.8% to 20.7%. All seven predictors retained their statistical significance after bootstrapping, confirming the validity of the score.

Table 1 Multivariate logistic regression analysis: independent predictors and the Pediatric Nosocomial Sepsis Prediction Score (Pe-NoSeP)

Figure 1

Receiver operator characteristic (ROC) curve for the prediction model.

Conclusion

The Pe-NoSeP score is a simple, easy-to-use bedside prediction model, which estimates the probability of NS in a child on a particular day and assists clinical decision-making. This tool may have diagnostic, therapeutic and preventive utilities based on its application.