Re-evaluating and recalibrating predictors of bacterial infection in children with cancer and febrile neutropenia

BACKGROUND: Numerous paediatric febrile neutropenia (FN) clinical decision rules (CDRs) have been derived. Validation studies show reduced performance in external settings. We evaluated the association between variables common across published FN CDRs and bacterial infection and recalibrated existing CDRs using these data. METHODS: Prospective data from the Australian-PICNICC study which enrolled 858 FN episodes in children with cancer were used. Variables shown to be significant predictors of infection or adverse outcome in >1 CDR were analysed using multivariable logistic regression. Recalibration included re-evaluation of beta-coefficients (logistic model) or recursive-partition analysis (tree-based models). FINDINGS: Twenty-five unique variables were identified across 17 FN CDRs. Fourteen were included in >1 CDR and 10 were analysed in our dataset. On univariate analysis, location, temperature, hypotension, rigors, severely unwell and decreasing platelets, white cell count, neutrophil count and monocyte count were significantly associated with bacterial infection. On multivariable analysis, decreasing platelets, increasing temperature and the appearance of being clinically unwell remained significantly associated. Five rules were recalibrated. Across all rules, recalibration increased the AUC-ROC and low-risk yield as compared to non-recalibrated data. For the SPOG-adverse event CDR, recalibration also increased sensitivity and specificity and external validation showed reproducibility. INTERPRETATION: Degree of marrow suppression (low platelets), features of inflammation (temperature) and clinical judgement (severely unwell) have been consistently shown to predict infection in children with FN. Recalibration of existing CDRs is a novel way to improve diagnostic performance of CDRs and maintain relevance over time. FUNDING: National Health and Medical Research Council Grant (APP1104527).