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

Abstract

Intestinal perforation (IP) in preterm infants is a life-threatening condition that may result in serious complications and increased mortality. Early Prediction of IP in infants is important, but challenging due to its multifactorial and complex nature of the disease. Thus, there are no reliable tools to predict IP in infants. In this study, we developed new machine learning (ML) models for predicting IP in very low birth weight (VLBW) infants and compared their performance to that of classic ML methods. We developed artificial neural networks (ANNs) using VLBW infant data from a nationwide cohort and prospective web-based registry. The new ANN models, which outperformed all other classic ML methods, showed an area under the receiver operating characteristic curve (AUROC) of 0.8832 for predicting IP associated with necrotizing enterocolitis (NEC-IP) and 0.8797 for spontaneous IP (SIP). We tested these algorithms using patient data from our institution, which were not included in the training dataset, and obtained an AUROC of 1.0000 for NEC-IP and 0.9364 for SIP. NEC-IP and SIP in VLBW infants can be predicted at an excellent performance level with these newly developed ML models. https://github.com/kdhRick2222/Early-Prediction-of-Intestinal-Perforation-in-Preterm-Infants .

Figure 1

Architecture of machine learning models. (a) Model 1 is a baseline neural network based on the conventional multilayer perceptron (MLP) architecture. (b) Model 2 is composed of two different MLPs. One branch predicts NEC, and the other branch predicts either NEC-IP or SIP. The feature vectors from the 3rd layer of the network for NEC are concatenated with the 4th layer of the MLP branch for NEC-IP and SIP. (c) Model 3 has the same network architecture as that of Model 1. Pretrained Model 1 for NEC was further fine-tuned to estimate NEC-IP/SIP.

Figure 2

Receiver operating characteristic curves of proposed ML models for (a) NEC prediction, (b) NEC-IP prediction, and (c) SIP prediction.

Figure 3

Receiver operating characteristic curves of proposed ML models from 57 test cases. (a) NEC prediction, (b) NEC-IP prediction, and (c) SIP prediction.

Figure 4

Flowchart of data processing. Input values were categorized into ordinal, continuous and categorical types. To solve the data imbalance problem, oversampling and undersampling technique were applied. Then, the data were normalized.