PREDICTING LATE-ONSET SEPSIS OUTCOME IN PRETERM INFANTS USING BIG DATA ANALYSIS

Background and aims Sepsis remains a significant cause of morbidity and mortality in preterm infants. Late-onset sepsis (LOS), defined as occurring after 72 of age, affects especially those born most premature. The use of invasive therapy, such as mechanical ventilation, venous catheterization, and prolonged parenteral nutrition, increases the risk. Antibiotic treatment is started after clinical symptoms. However, these symptoms can be difficult to recognize. Blood culture remains the best available diagnostic tool. However, the definitive diagnosis requires up to 48 hours and may produce false negative and false positive results. This study aimed to build a model to accurately predict the outcome of a blood culture drawn for LOS using advanced analytics of medical records. Methods All infants born <32 w between April 2008 and September 2017 admitted to the NICU of the UMC Utrecht within 24 hours after birth were included in the study if at least one blood culture was taken after 72 hours after birth and antibiotics were started. All available data from electronic medical records were anonymized and made available for analysis. The data up to three days before the onset (blood culture) of LOS were considered. Using Gradient Boosting, 70% of the data was studied as training set and 30% of the data as test set. Subsequently, a best fit model was identified. ROC curves with corresponding area under the curves (AUCs), sensitivity, specificity, positive predictive value and negative predictive value were produced. Results In total 2379 infants were evaluated of which 620 had at least one suspicion of LOS (26%). The LOS-infants were younger (29.9 ± 2.4 vs. 28.5 ± 2.6; p< 0.01) and had a lower birth weight (1343 ± 402gr vs. 1103 ± 360gr; p<0.01). 1198 blood cultures were available for model building of which 442 (37%) had a positive (culture positive) result and 756 (63%) a negative result. The combination of clinical variables recorded in electronic patient records allowed for the discrimination between positive and negative blood cultures (area under the curve, p-value, sensitivity, specificity, negative predictive value, 0.85, p<0.01, 0.86, 0.58, 0.88). Conclusions In preterm infants, the outcome of LOS could be predicted with the clinical characteristics of a preterm infant at the clinical onset of LOS with great accuracy using advanced analytics of electronic medical records.