A Prognostic Score for Predicting 30-Day Mortality After Gram negative Bloodstream Infection in Pediatric Patients with Hematological Malignancy

Abstract

Abstract Background Patients with hematological malignancies (HMs) are at high risk of gram-negative bloodstream infections (GN-BSI) due to prolonged immunosuppression and neutropenia. These infections are associated with significant morbidity and mortality, often exceeding 30%. Early identification of high-risk patients is crucial, yet reliable prognostic tools remain limited. This study aimed to develop a validated, practical, and sensitive model for predicting 30-day mortality following GN-BSI in pediatric patients with HMs. Methods We analyzed 1,299 GN-BSI episodes from 867 patients. The dataset was randomly partitioned into training (70%, n=881 episodes, 173 deaths) and validation (30%, n=406 episodes, 82 deaths) cohorts. Univariate logistic regression screening identified 12 significant predictors (p<0.05) from 32 candidate variables. A multivariable logistic regression model was developed including all univariate-significant predictors without stepwise selection to ensure transparency and reproducibility. Model performance was assessed using the C-statistic (AUC), Brier score, calibration metrics, and bootstrap internal validation (1,000 replicates). Results The final 12-predictor model demonstrated strong performance in both training (AUC = 0.934, 95% CI 0.913–0.955) and validation (AUC = 0.936, 95% CI 0.912–0.961) datasets, with good calibration (Hosmer–Lemeshow p = 0.726) and minimal overfitting (bootstrap optimism = 0.0007). The strongest predictors for 30-day mortality were need for inotropic support (OR = 57.4, 95% CI 32.0–102.8, p < 0.0001), Acinetobacter infection (OR = 4.2, 95% CI 1.6–11.0, p = 0.003), low platelet count (OR = 0.71 per SD decrease, 95% CI 0.54–0.93, p = 0.014), and extremely drug-resistant organisms (OR = 2.7, 95% CI 1.1–6.6, p = 0.026), while E. coli infection, was associated with markedly reduced odds of 30-day mortality (OR = 0.45, 95% CI 0.22–0.93, p = 0.032). To address the time-varying predictor limitation (inotropic support unavailable at initial presentation), we developed two complementary tools: (1) an 11-predictor Early Risk Nomogram (validation AUC = 0.728) and (2) a simplified integer scoring system using the Sullivan method (AUC = 0.709). The integer score showed only 2.7% discrimination loss compared with the nomogram but reduced calculation time to <1 minute, enabling rapid bedside use and early risk stratification in resource-limited settings. Conclusion We developed a reliable, well-calibrated 30-day mortality prediction model for pediatric GN-BSI. The early-risk tools enable rapid bedside risk calculation at initial BSI presentation, supporting empirical antibiotic selection, monitoring intensity decisions, and family counseling, practically in resource-limited settings. External prospective validation is needed before clinical implementation.