Computational Approaches to Penicillin Allergy De-labelling: A Clinical Review

Penicillin allergy is recorded in roughly one in ten hospital patients, yet fewer than one in ten of those records reflects a real, clinically significant allergy. The mismatch matters: patients with a penicillin label receive more second-line and broader-spectrum antibiotics, with measurable downstream costs in infection, surgical site morbidity, length of stay, and antimicrobial resistance. Specialist-led skin testing and graded oral challenge remain the diagnostic standard, but the global allergy workforce cannot test the population that needs testing. Validated decision rules such as PEN-FAST, structured non-allergist pathways, and direct oral challenge in low-risk patients have closed part of that gap, but uptake is still limited by inconsistent free-text documentation, variable clinician confidence, and weak integration with electronic health records. This review summarises how computational methods are being added to that pathway. Supervised machine-learning models can mine electronic health records to identify candidates for review. Natural language processing can recover the timing, character, and severity of a reaction from clinic letters that the structured allergy field omitted. Predictive models that combine clinical, laboratory, and immunological inputs are starting to outperform rule-based scores on beta-lactam risk. Decision-support systems are moving past simple drug-name matching toward context-aware alerts. Image classifiers and large language models are opening a route to rapid, photograph-and-history triage. We discuss what is needed for any of this to be safe in routine care: representative training data, transparent model behaviour, regulatory alignment, and prospective evaluation against the outcomes the antimicrobial stewardship community already measures.