A custom GPT-based model for the automated analysis and interpretation of antimicrobial susceptibility tests in Gram-negative bacteria.

INTRODUCTION: Antimicrobial resistance in Gram-negative bacilli represents a growing clinical challenge that demands innovative tools to optimize antibiogram interpretation and support safe therapeutic decisions. OBJECTIVE: To develop and validate a custom generative pre-trained transformer (GPT) model for antibiogram interpretation, comparing its performance with an alternative model based on Gemini (Gems). METHODS: A retrospective, cross-sectional diagnostic accuracy study was conducted following the STARD-artificial intelligence (AI) guideline. A total of 210 antibiograms were analysed, balanced across controls and five resistance mechanisms (penicillinases, broad-spectrum β-lactamases, extended-spectrum β-lactamases, AmpC and Carba-R). The custom GPT, configured through prompt engineering and retrieval-augmented generation, was evaluated against expert interpretation and compared with Gems. Performance metrics (sensitivity, specificity, precision, F1-score, receiver operating characteristic/precision-recall area under the curve and Cohen's Kappa) were calculated, and errors were classified by clinical impact as overtreatment or undertreatment. RESULTS: The GPT achieved an accuracy of 95.7% (95% confidence interval: 92.0-98.0) and a Kappa of 0.948, outperforming Gems (91.9% and 0.902, respectively). Both models showed good performance, but GPT displayed greater consistency across metrics, with sensitivities and F1-scores >90% in all classes. Clinically relevant errors were fewer and less severe with GPT (9 versus 17 in Gems), with markedly lower spectrum penalty and treatment failure risk indices. CONCLUSIONS: This study provides proof-of-concept evidence that a custom GPT model can accurately interpret antibiograms of Gram-negative bacilli under controlled retrospective condition. Although these findings are promising, external multicentre validation, expansion to other pathogens and resistance mechanisms and further assessment as a clinical decision-support tool are required before routine clinical implementation.