Artificial intelligence for predicting 30-day mortality after emergency laparotomy

Purpose Accurate pre-operative risk prediction in emergency laparotomy (EL) is essential for appropriate resource allocation and improve patient outcomes. This study aimed to establish the accuracy of an artificial intelligence (AI) model in predicting 30-day mortality after EL using a national dataset. Methods Data was extracted from the Australian and New Zealand Emergency Laparotomy Audit-Quality Improvement (ANZELA-QI) database from 1 July 2018 to 24 July 2023. AI models were created employing multilayer perceptron (MLP) and radial basis function (RBF) architectures. Extended AI models were developed using all available pre-operative variables and simplified AI models were developed using statistically significant variables identified through univariate analysis. We assessed the performance of AI models with that of a multivariate logistic regression (LR) and the UK-based National Emergency Laparotomy Audit (NELA) models using the area under the receiver operator characteristic curve (AUROC). Results Data from 8293 ELs were included in the model and were randomly divided into a training dataset of 6619 (80%) and a testing dataset of 1674 (20%). There were 537 (6.5%) deaths 30 days after EL. In the testing dataset, the NELA model performed the best (AUROC 0.836), followed by the multivariate LR model (0.824), MLP simplified (0.817). MLP extended (0.802), RBF extended (0.719), and RBF simplified (0.672). Pairwise comparisons showed no significant difference in discrimination among the NELA, multivariate LR and MLP models. Conclusion Pre-operative MLP and multivariate LR models demonstrated comparable performance to the NELA model in predicting 30-day mortality after EL.