A comparative analysis of logistic regression (LR) and artificial neural network (ANN) models for predicting antimicrobial resistance in surgical ICU patients: Insights from real-world evidence in India

Abstract

Background

Machine learning approaches for the prediction of antimicrobial resistance (AMR) are gaining attention but are yet to be commonly applied in practice.

Objective

This study aims to predict the AMR in surgical intensive care unit patients using logistic regression (LR) and artificial neural network (ANN) model.

Methods

Surgical ICU patients with resistant infections, regardless of the microorganism, were considered cases. Those with susceptible or no infections were considered controls. A total of 104 variables for patient characteristics, disease-related and clinical parameters, and surgical, culture, and prescription details were tested for the prediction of AMR using two methods: LR and ANN. The dataset was divided into a training (n = 3179) and a test (n = 1363) set. The outcome was considered a binary outcome: resistant infection and sensitive infection. Model evaluation metrics were an area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Predictive analysis was performed by using R software.

Results

Out of 8010 ICU patients, 4542 patients underwent surgery. Out of these surgical ICU patients, 36.90% were cases and 63.09% were controls. Both models performed similarly concerning sensitivity (ANN 86.6%; LR 86%), while improvement was found with respect to accuracy (ANN 88.2%; LR 86%), specificity (ANN 91.2%; LR 86%), AUROC (ANN 94%; LR 93%), and NPV (ANN 82.8%; LR 91%).

Conclusions

The ANN model has more predicting performance than the LR model to predict AMR in surgical ICU patients. These prediction algorithms may assist clinical decisions to aid the prevention of AMR.