EUVIMEDEUVIMED entdecken
CrossrefInfokarte

The accuracy and utility of large language models for extracting stewardship data from handwritten and scanned-in records

Rory D. Ostrow, Samuel Golenbock, Jacquelyn Pendergast, Hillary J. Mull, Westyn Branch-Elliman

Antimicrobial Stewardship & Healthcare Epidemiology · 2026 · 6 (1)

DOI: 10.1017/ash.2026.10919 ↗
AbstractBibliografische Angaben

Abstract

Abstract Background: Some clinical documentations, including surgical antibiotic administration data and hand-written notes, are only captured in non-digitized scanned documents, limiting their accessibility. Commercially available artificial intelligence (AI) tools may offer a means of extracting such data from scanned-in images, however, little is known about their accuracy for this application. Thus, the objective of this study was to evaluate the performance of two commercially available, protected AI tools for extracting antibiotic administration data as part of a larger study evaluating the sustainability of stewardship goals achieved during the Surgical Care Improvement Project mandatory reporting period. Methods: A sample of surgical cases with antibiotic administration data documented only in non-digitized scanned records was manually reviewed and submitted to two commercially available AI tools (Claude Sonnet version 4.6 and OpenAI version 5.0). The AI tools were prompted to extract data about antibiotic name, dose, and administration time. Results: The total accuracy across all three metrics (drug, dose, timing of administration) was limited (3/14, 21.4% for Claude and 2/14, 14.3% for OpenAI), but overall excellent for identifying that an antibiotic had been administered and the type of drug administered (Claude, 100%, OpenAI, 92.8%). AI tools struggled to identify total dose administered (Claude, 50%, OpenAI, 42.9%), primarily due to challenges distinguishing aliquots from administered doses. Accuracy for identifying timing was also variable, particularly when administration time was documented using a graphic scale rather than written text. Conclusions: Potential utility of these commercial tools for extracting antimicrobial use data from scanned-in records depends upon review goals and what level of precision is necessary.

Wissen, das zu Ihrer Arbeit passt.

Vom kostenlosen Einstieg über die persönliche Recherche bis zur vollständigen Research Suite.

Tarife ansehen →

Sind alle Verlagsvolltexte enthalten?

Nein. Bibliografische Nachweise, Abstracts und Volltexte sind unterschiedliche Inhalte. Ob ein Volltext verfügbar ist, hängt von der Quelle, Open Access und Ihren vorhandenen Lizenzen ab. Ein EUVIMED-Tarif ersetzt keine zusätzlichen Verlagsrechte.

Tarife ansehen →