AB191. SOH26AB_0456. Pilot testing of an AI-powered prescribing support tool for identifying drug-drug interactions in surgical inpatients

Background: Polypharmacy significantly increases the risk of drug-drug interactions (DDIs) in hospitalised surgical patients, necessitating rapid and reliable detection to enhance medication safety. Artificial intelligence (AI) tools may support clinicians by improving the speed and accuracy of DDI identification. This pilot study aimed to evaluate a prototype AI-powered prescribing assistant and compared its performance with clinical pharmacist review and a commonly used large language model (ChatGPT).

Methods: Nineteen surgical inpatient medication lists were analysed. The British National Formulary (BNF) served as the gold-standard reference for all DDIs. Three modalities assessed each case: pharmacist review, the prototype AI tool (Katana AI), and ChatGPT. For each source, the total number of flagged interactions, confirmation against the BNF, time to generate results, and severity classification (moderate, cautionary, severe) were recorded. All findings were independently verified by clinicians and pharmacists.

Results: Across 19 cases, 156 medications were reviewed (median, 8; range, 3–14). Katana AI identified 49 DDIs, of which 47 were confirmed, achieving 97.9% accuracy. Pharmacists flagged 81 DDIs with 73 confirmed (90.1% accuracy), while ChatGPT flagged 142 DDIs, with 92 confirmed (64.8% accuracy). The AI tool generated reports in a mean time of 44 seconds, similar to ChatGPT (41 seconds) and markedly faster than pharmacists (249 seconds). A minor software limitation accounted for two missed interactions.

Conclusions: This pilot study demonstrates that a prototype AI-powered prescribing assistant can accurately and rapidly identify clinically relevant DDIs in surgical inpatients. The tool outperformed ChatGPT and closely matched pharmacist accuracy, while delivering results in a fraction of the time required for manual review. Although a minor software limitation resulted in a small number of missed interactions, these early findings highlight the potential of AI systems to strengthen medication safety, reduce human-error risk, and enhance prescribing efficiency in acute care settings. Continued refinement and larger-scale clinical validation are warranted before routine clinical integration.

Keywords: Artificial intelligence (AI); drug-drug interactions (DDIs); medication safety; polypharmacy; prescribing support tools