JMIR Publications

ABSTRACT

Background:

Necrotizing enterocolitis (NEC) is the most common gastrointestinal emergency affecting preterm infants with high mortality and morbidity. With suboptimal and incomplete methods of prevention of NEC, early diagnosis and treatment can potentially mitigate the impact of NEC. This study explores the application of machine learning techniques, specifically Random Forest and Extreme Gradient Boosting (XG Boost), to improve early and accurate NEC and FIP diagnosis.

Objective:

To evaluate the effectiveness of sampling techniques in addressing class imbalance and to identify the optimal machine learning (ML) classifiers for predicting necrotizing enterocolitis (NEC) and focal intestinal perforation (FIP) in preterm infants.

Methods:

We developed ML models using 49 clinical variables from a retrospective cohort of 3,463 preterm infants, using clinical data from the first two weeks of life as input features. We applied various sampling strategies to address the inherent class imbalance, and then combined various sampling strategies with different ML algorithms. Parsimonious models with selected key predictors were evaluated to maintain predictive performance comparable to the full-featured (complex) models.

Results:

The parsimonious generalized linear model (GLM) with SMOTE sampling achieved an area under the receiver operating characteristic curve (AUROC) of 0.79 for NEC prediction, closely approximating the complex model's AUROC of 0.76. For FIP prediction, parsimonious models of GLM with ADASYN sampling and XG Boost with TOMEK sampling achieved AUROC values exceeding 0.90, comparable to those of the corresponding complex models. For both NEC and FIP, the area under the precision-recall curve (AUPRC) surpassed the respective prevalence rates, indicating strong performance in identifying rare outcomes.

Conclusions:

We demonstrate that targeted sampling strategies can effectively mitigate class imbalance in neonatal datasets, and simplified models with fewer variables can offer comparable predictive power, enhancing the performance of ML-based prediction models for NEC and FIP.