Supported via arm's length funding from MSD.

Highlights from

The European Congress of Clinical Microbiology & Infectious Diseases

29th Annual Meeting

Amsterdam 13-16 April 2019

Neural networks can improve empirical antibiotic prescribing

Take-home messages
  • An artificial neural network is a series of computer algorithms loosely modelled on biological neuronal networks
  • Neural networks can be used help to clinicians make clinical decisions
  • It was possible to predict with 98% accuracy whether patients with high-risk febrile neutropenia would develop MDR-GNB infection using neural networks
"Neural networks are helpful in predicting which patients with high-risk febrile neutropenia will have multidrug-resistant gram-negative bacterial infection."

Dr Carolina Garcia Vidal, Infectious Diseases Consultant, Hospital Clínic de Barcelona, Spain

Neural networks can be used to identify patients at risk of developing different types of bacterial infection, including multidrug-resistant (MDR) strains, according to new research presented at the European Congress of Clinical Microbiology and Infectious Diseases (ECCMID) 2019, Amsterdam, the Netherlands. This could help clinicians improve management of patients and guide empirical antibiotic prescribing, explained Dr Carolina Garcia-Vidal, Infectious Diseases Consultant, Hospital Clínic de Barcelona, Spain.

In one study, Dr Garcia-Vidal and her team predicted which patients with high-risk febrile neutropenia would develop MDR-gram-negative-bacterial (GNB) infection with an accuracy of 98% using neural networks.

“If you come to my hospital now, and you have a febrile neutropenia episode at the onset, I can predict that you are going to have a multidrug resistant gram-negative bacterial infection with 98% accuracy rate,” said Dr Garcia-Vidal. ”We have done that by analysing more than 440 million pieces of data.”

Prescribing appropriate empirical antibiotic treatment to patients is notoriously challenging. “In a recent study of 3,235 febrile neutropenia episodes, 46% received inappropriate empirical antibiotic treatment,” said Dr Garcia-Vidal.

An artificial neural network is a series of computer algorithms, loosely based on a neuronal network in a living brain, that is designed to recognise patterns in data. Artificial neural networks can be trained to solve problems by training the model with sample data.

A biological neuron has multiple unique inputs and produces one output. This is also the case for artificial neural networks, in which a range of different inputs are processed and modified by a mathematical equation called a weight.

An artificial neural network combines different weighted inputs together to produce a final value. Creating an artificial neural network involves a training stage and a test stage.

Dr Garcia-Vidal and her team extracted data from 7 million electronic health records (EHRs). Patients were randomly selected for the learning stage (70% of the cohort, training set), as well as for the test stage (30%).

A neural network consisting of 14 input parameters previously selected by multivariate analyses was used. It was then determined whether the parameters were significant or not for the network. Parameters that were not significant were eliminated, and the network was trained once again.

In this study, 3,235 episodes of high-risk febrile neutropenia (HRFN) in haematological patients were documented (median age: 57 years; interquartile range: 44-67). The data included 56.9% males and 43.1% females. 38% of patients had acute leukaemia and 28% had received human stem cell transplantation.

Infections caused by MDR-GNB accounted for 180 (5.6%) episodes. The most frequent MDR-GNBs were due to Pseudomonas aeruginosa (53%) and extended-spectrum beta-lactamase-GNB (46%).

The model predicted with an accuracy of 98% that 236 patients (7%) would have MDR-GNB infection (2% were false positives) and 2,999 would not have MDR-GNG infection (3% of patients presenting with MDR-GNB infection were false negatives).

In another study presented at ECCMID, Dr Garcia-Vidal showed how neural networks predicted the number of infections caused by P. aeruginosa in haematological patients with high-risk neutropenia with an accuracy of 96%.

“This tool might be used to analyse data from EHRs in real time and provide a revolutionary approach that can be used as a decision-support system for optimising microbiological testing [and providing] faster diagnoses and antibiotic treatments,” said Dr Garcia-Vidal.

According to Dr Garcia-Vidal, ability to use neural networks to guide clinical decision-making has been enabled by two revolutions. The first is the vast amount of data that can be mined from electronic health records. The second revolution is the onset of machine learning and neural networks, enabled by supercomputers that can provide results in real time.

Based on:

Garcia-Vidal C. Applying artificial intelligence to improve empiric antibiotic treatment (symposium S0620). Presented on Tuesday 16 April 2019

Garcia-Vidal C, Puerta P et al. Predicting multidrug-resistant Gram-negative infections in haematological patients with high-risk febrile neutropenia using neural networks (oral presentation O1185). Presented on Tuesday 16 April 2019.

Garcia Vidal, Sanjuan G et al. Predicting infections caused by Pseudomonas aeruginosa in haematological patients with high risk neutropenia via the use of neuronal networks (poster P2438).

Top image: © iStockPhoto nicolas_

Article image: iStockPhoto Svisio

The content and interpretation of these conference highlights are the views and comments of the speakers/authors.

Supported by