Performance comparison of prediction models for neonatal sepsis using logistic regression, multiple discriminant analysis and artificial neural network

Abstract

Sepsis is one of the major causes of neonatal deaths worldwide. Majority of these neonatal deaths occurs in resource-poor countries due to inaccessibility of hospitals and absence of laboratories. Blood culture which is the gold standard to confirm sepsis is time-consuming and requires the presence of a laboratory. To start the antimicrobial therapy at the earliest, prediction models have been developed. A vast number of available prediction models require laboratory tests and cannot be used in the developing countries where such facilities do not exist. Therefore, there is a need for non-invasive prediction models. The objectives of this study are as follows: - (i) to train and test non-invasive prediction models for neonatal sepsis (ii) to compare the performance of the invasive with non-invasive prediction models. For this retrospective study, we extracted the data of 1446 neonates from the Medical Information Mart for Intensive Care (MIMIC) III data set. We trained and tested six prediction models using this data set. Three of these six models were trained using non-invasive parameters (model LR(NI), model ANN(NI) and model MDA(NI)) and three were trained using invasive and non-invasive parameters (model LR(O), model ANN(O) and model MDA(O)). The sensitivity of model LR(NI), model ANN(NI), model MDA(NI), model LR(O), model ANN(O) and model MDA(O) at their optimum threshold values were 81.68%, 79.39%, 82.44%, 77.10%, 79.39% and 78.63% respectively. Whereas, specificity of the above mentioned models were 82.27%, 81.82%, 80.00%, 84.77%, 82.05% and 78.30% respectively. To decrease the neonatal mortality rate in resource-poor areas one may use non-invasive prediction models where invasive parameters are not available due to lack of resources, as shown by our study that non-invasive prediction models can achieve similar predictive capability as the invasive prediction models.