Artificial Intelligence and Machine Learning Applications in Antimicrobial Resistance Research: A Systematic Review

Abstract
Background
Antimicrobial resistance (AMR) represents an existential global health crisis, causing over one million deaths yearly all over the world. Conventional culture-based susceptibility testing requires 48–72 hours, an interval insufficient for guiding empirical therapy in critically ill patients. Artificial intelligence (AI) and machine learning (ML) approaches have emerged as probable alternative solutions, capable of analysing high-dimensional clinical, spectral, and genomic datasets to generate rapid resistance predictions.
Objectives
To systematically identify, appraise, and synthesise peer-reviewed evidence on AI/ML model applications for detecting, predicting, and managing AMR in bacterial pathogens using human clinical data.
Methods
A systematic search of PubMed, Scopus, and Embase was conducted for studies published between January 2019 and January 2026. Eligible studies incorporated supervised ML or deep learning models applied directly to AMR phenotype or genotype prediction using human clinical data, with quantitative performance metrics. Data extraction followed the CHARMS checklist; methodological quality was assessed using PROBAST. PRISMA 2020 guidelines were followed throughout.
Results
From 4,067 records identified, 57 studies met full eligibility criteria (after removal of 452 duplicates and exclusion of 3,554 records at screening and full-text review stages). The predominant algorithms were Random Forest (n ≈ 22), XGBoost (n ≈ 20), LightGBM (n ≈ 14), logistic regression with regularisation (n ≈ 18), and support vector machines (n ≈ 16). MALDI-TOF mass spectrometry spectral data and whole-genome sequencing features were the most commonly employed input modalities. Best-performing models achieved AUROC values of 0.85–0.99 for pathogens including MRSA, carbapenem-resistant Klebsiella pneumoniae (CRKP), and Acinetobacter baumannii. PROBAST assessment revealed high risk of bias in the Participants and Analysis domains across all 57 studies, predominantly due to retrospective single-centre designs and absent external validation.
Conclusions
Tree-based ensemble methods, particularly XGBoost and LightGBM applied to MALDI-TOF spectral inputs, demonstrate strong diagnostic potential for rapid AMR prediction. However, pervasive methodological limitations including retrospective designs, absent external validation in ~ 65% of studies, and unreported missing data handling preclude immediate clinical translation. Prospective multicentre validation, explainability integration, and standardised reporting frameworks are required before clinical adoption.