Development of artificial neural networks for early prediction of intestinal perforation in preterm infants

Area of Science:

• Neonatal Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Intestinal perforation (IP) poses a severe threat to preterm infants, leading to complications and mortality.
• Early prediction of IP is crucial but difficult due to the condition's complexity.
• Current tools for predicting IP in infants lack reliability.

Purpose of the Study:

• To develop and evaluate novel machine learning (ML) models for predicting IP in very low birth weight (VLBW) infants.
• To compare the performance of new artificial neural network (ANN) models against traditional ML methods.
• To establish reliable predictive tools for IP in VLBW infants.

Main Methods:

• Development of artificial neural networks (ANNs) using nationwide cohort and registry data of VLBW infants.
• Training and validation of ML models, including ANNs and classic ML methods.
• Independent testing of developed algorithms on institutional patient data not used in training.

Main Results:

• The newly developed ANN models demonstrated superior performance compared to classic ML methods.
• Achieved an area under the receiver operating characteristic curve (AUROC) of 0.8832 for predicting necrotizing enterocolitis-associated IP (NEC-IP).
• Attained an AUROC of 0.8797 for spontaneous IP (SIP) and 1.0000 (NEC-IP) and 0.9364 (SIP) on independent test data.

Conclusions:

• Newly developed ML models, specifically ANNs, can predict NEC-IP and SIP in VLBW infants with excellent accuracy.
• These models offer a promising advancement in the early detection of intestinal perforation in preterm infants.
• The findings suggest a potential for improved clinical management and outcomes for VLBW infants at risk of IP.