JMIR Publications

ABSTRACT

Background:

Neonatal sepsis is associated with most mortalities and morbidities in the neonatal intensive care unit (NICU). Many studies have developed prediction models for early diagnosis of bloodstream infection in newborns, but there are limitations to data collection and management because such models are based on high-resolution waveform data.

Objective:

This study aimed to check the feasibility of a prediction model using non-invasive vital signs data and machine learning technology.

Methods:

This study used the Medical Information Mart for Intensive Care III clinical database, which published ICU electronic medical record data. The late-onset neonatal sepsis (LONS) prediction algorithm was based on NICU inpatient data and designed to detect clinical sepsis 48 hours before occurrence. The performance of this prediction model was evaluated using various feature selection algorithms and machine learning models.

Results:

The LONS prediction model was found to have performance comparable to prediction models that use invasive data such as high-resolution vital signs, blood gas estimations, blood cell counts, and pH levels from previous studies. The area under the receiver operating characteristics (AUROC) of the 48-hour prediction model was at 0.861, and the AUROC of the onset detection model was 0.868. Main features that could be a candidate vital marker for clinical neonatal sepsis were blood pressure, oxygen saturation, and body temperature. Feature generation using kurtosis and skewness of the features showed the highest performance.

Conclusions:

The results of this study confirmed that the LONS prediction model based on machine learning can be developed using vital signs data that are regularly measured clinically. Future studies require external validation using different types of data sets and actual clinical verification of the developed model.