Effect of an Artificial Intelligence-assisted Antibiotic Susceptibility Test on Reducing the Mortality Rate of and Antibiotics Use in Patients with Bacteremia: A Prospective Observational Cohort Study

Abstract

No mass spectrometry-based antibiotic susceptibility test (AST) AI system has been robustly validated and truly deployed in real-world settings. In the study, an AI-assisted AST prediction system XBugHunter was extensively validated and deployed. A total of 90,064 consecutive patients were included in the models development. During the deployment, 155 consecutive patients with Staphylococcus aureus bacteremia (SAB) was analyzed, whereas the historical control cohort included 455 patients. The predictive sensitivity and specificity for S. aureus (oxacillin) were 0.95 and 0.97, respectively. The reduction on AST turnaround time was 34.06 h. Death within 28 days was 13.55% for the SAB patients in the XBugHunter intervention group, which was significantly lower than 17.58% in the historical control cohort. Totally, 114.77 defined daily doses of inappropriate antibiotics were avoided by deploying XBugHunter. This work showed that XBugHunter can prevent inappropriate antibiotic use, and the adjustment in antibiotic treatment can yield a lower mortality rate.

Rapid and Accurate Discrimination of Mycobacterium abscessus Subspecies Based on Matrix-Assisted Laser Desorption Ionization-Time of Flight Spectrum and Machine Learning Algorithms

“…At the same time, the AUROC of the ML model attained 0.89 [33]. In another large-scale research of ESKAPE pathogens, all the t-SNE plots of MALDI-TOF spectra showed no specific pattern for discriminating resistant pathogens from susceptible ones [22]. Likewise, the predictive performance attained as high as 0.95 for detecting oxacillin resistance for Staphylococcus aureus.…”

“…The peak is suspected to be in accordance with the peak of m/z 9478 that was found in our study (Figure 9B). The subtle difference between m/z 9473 and 9478 resulted from shifting problem caused by isotopes, so m/z 9473 and 9478 could be regarded as the same peptide molecules [22,37]. Based on previous studies, peak of m/z 9478 occurred less than 40% in either subsp.…”

“…Machine learning (ML) technology has been able to save considerable time in analyzing complex data without human error [19,20]. ML-based methods have been reported to perform rapid and accurate detection of resistant bacteria (e.g., E. coli [21], S. aureus, E. faecium [22]) from the MALDI-TOF MS spectrum. Additionally, subspecies identification can be achieved by applying ML to analyze the MALDI-TOF MS spectrum.…”