Artificial Intelligence for Antimicrobial Resistance Detection and Prediction in Klebsiella pneumoniae: A Systematic Review of Clinical Microbiology Applications.

内容摘要

Klebsiella pneumoniae is a WHO critical-priority pathogen associated with a substantial antimicrobial resistance (AMR) burden. Conventional microbiology workflows, including antimicrobial susceptibility testing, often require 36-72 hours, prolonging empirical therapy and contributing to antibiotic overuse. Artificial intelligence (AI) has emerged as a promising approach for enhancing the detection and prediction of antimicrobial resistance. We searched four databases (PubMed, EMBASE, MEDLINE, and CENTRAL) from 1 January 2010 to 3 January 2026 for peer-reviewed, original research studies evaluating AI methods for the detection and/or prediction of AMR in K. pneumoniae. Studies without K. pneumoniae-specific extractable outcomes were excluded. Data on study characteristics, input modalities, AI methods, performance, workflow gains, and validation methods were extracted and narratively synthesised. Risk of bias was assessed using PROBAST and QUADAS-2 according to study design. Fifty-seven studies were included, with publication output accelerating sharply in 2024-2025 (27/57, 47.4%). Most studies originated from East Asia and predominantly aimed to classify resistance phenotypes from pre-AST data (37/57, 64.9%) using machine learning approaches. MALDI-TOF mass spectrometry was the most common input modality (27/57, 47.4%), followed by genomic sequencing and vibrational spectroscopy (12/57 each, 21.1%). Random forests were the most frequently studied model family (28/57, 49.1%), with high reported discrimination. Among AUROC/AUC-primary studies, 26/35 (74.3%) reported best-model performance ≥0.90; however, overall risk of bias was high, present in 45/57 studies (78.9%). Internally validated study designs predominated, with external validation reported in only 17/57 studies (29.8%), and prospective, real-world evaluation in 1/57. AI-based AMR prediction and detection in K. pneumoniae is advancing rapidly, with MALDI-TOF-enabled approaches appearing most readily translatable to clinical microbiology workflows. However, the field remains dominated by retrospective, internally validated studies, often using imperfect automated susceptibility systems as reference standards. Progress now depends on rigorous external and prospective multicentre validation using geographically diverse datasets. Klebsiella pneumoniae is a bacterium that can cause serious infections, especially in people admitted to hospital. Many strains are now resistant to antibiotics, which means some medicines no longer work well. Standard laboratory tests to find the right antibiotic usually take 36 to 72 hours. During this time, doctors often need to start treatment with broad-spectrum antibiotics before the full results are ready, which can contribute to antibiotic resistance. We reviewed published studies to understand whether artificial intelligence (AI) could help identify antibiotic resistance in Klebsiella pneumoniae more quickly. AI is a type of computer system that can look for patterns in complex data. We found 57 studies published between 2018 and 2025. Most looked at whether AI could use information already produced in the microbiology laboratory, especially from a common test called “MALDI-TOF mass spectrometry” (a test that identifies bacteria from their protein “fingerprint”), to predict resistance earlier than standard testing. Studies predominantly originated from East Asia. Most reported good accuracy, and almost half said AI could provide faster results than usual laboratory methods. This means AI could help doctors choose better antibiotics sooner and reduce unnecessary use of broad-spectrum treatment. However, we identified some important limitations. Most studies tested the AI on past data rather than in real-time patient care. In some studies, the AI learned from automated systems rather than the best available (“gold standard”) methods, meaning it may reproduce their weaknesses. More real-world studies across different hospitals are needed before AI can be used safely.