Improving ICU Infection Predictions

A new machine learning model, called multi-objective symbolic regression (MOSR), shows promise in predicting bloodstream infections (BSI) and antimicrobial resistance (AMR) in ICU patients. Developed by researchers from King’s College London and Guy’s and St. Thomas’ NHS Trust, the model uses data available at ICU admission to guide timely infection treatment decisions.

Traditional machine learning methods have struggled with the high data imbalance typically seen in ICU infection cases. However, MOSR, which produces interpretable equations, can work with imbalanced datasets and achieve high prediction accuracy. In this study, MOSR consistently outperformed conventional models, with F1-Scores 0.28 points higher than the best-performing alternatives​.

The study involved 1,142 ICU patients and used 25 factors related to infection risk, including mean arterial pressure, bicarbonate, urea, and white blood cell count. The MOSR model prioritized sensitivity to minimize false negatives, which is critical for ICU infection control. MOSR demonstrated an AUROC score of 0.79 for BSI prediction and 0.86 for AMR, surpassing all other models tested. In leveraging routinely available patient data, MOSR provides a scalable tool that can be implemented across ICUs without the need for deep learning infrastructure.

“An important way to tackle the grave threat of antimicrobial resistance is to protect the antibiotics we already have, which goes hand in hand with the urgent need for fast diagnostics,” explains Lindsey Edwards, expert in microbiology at King’s College London, in the press release. “Using AI to speed up the diagnostics of infection [...] could not only have a huge impact on the patient’s survival and their care outcomes; but could help to preserve the antibiotics we already have developed and prevent the development of further antibiotic resistance."

This research has paved the way for further ongoing studies to validate MOSR’s effectiveness across ICU’s. The researchers hope more testing could fulfill the regulatory requirements for deployment of this AI approach across the NHS.