Validation, Calibration and Transparency, Not Accuracy: An Appraisal of Artificial Intelligence Against Antimicrobial Resistance and a Sustainable, Testable Diagnostic Architecture

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Artificial intelligence is widely proposed against antimicrobial resistance, but how much of that literature could support clinical deployment is unclear. We searched PubMed (2016–2026) across four domains — genomic prediction, clinical risk models, therapy, and rapid diagnostics — retrieving all 2,446 records. Of 725 eligible studies, 80 were appraised at full text against the artificial-intelligence extensions of the standard risk-of-bias and reporting tools (PROBAST+AI, TRIPOD+AI). Selection favoured externally validated work, so the proportions are upper bounds. Output grew from 10 studies in 2016 to 142 in 2025, with no evidence that genuinely external evaluation became more likely (odds ratio 1.08/year, 95% CI 0.89–1.34). Of the 80, 45% were externally validated, 24% released a model a reader could apply, 23% assessed calibration, 16% justified a sample size, and five compared performance across patient characteristics against eight across sites; 9% called a random split of their own data external validation. Risk of bias was low in 16, unclear in 55 and high in 9, and none of the 18 deep-learning studies reported calibration (95% CI 0–17.6%). Validation, calibration and transparency limit this field, not accuracy. We derive the design constraints these impose on a culture-independent diagnostic platform, a projection with no performance data.

This publication has 141 references indexed in Scilit: