5PSQ-148 Can artificial intelligence assist in identifying drug interactions in critically ill patients? ChatGPT vs standard clinical practice

Abstract

<h3>Background and Importance</h3> Artificial intelligence (AI) is progressively being integrated into daily clinical practice, including hospital pharmacy. However, rigorous validation is necessary to guarantee its reliability and clinical applicability. <h3>Aim and Objectives</h3> To evaluate the capability of a generative AI model to detect drug-drug interactions (DDIs) compared to databases commonly used in standard clinical practice (SCP). <h3>Material and Methods</h3> This observational retrospective study included hospitalised patients in critical care units (ICU and PACU) with ≥5 prescribed drugs on admission during January–February 2025. DDIs were analysed using two methods: AI model: Customised ChatGPT–4.0 analysed screenshots of prescribed medications and answered classification queries. Reference method: Manual consultation of UpToDate and Medscape databases. Interactions were classified according to LEXICOMP: contraindicated (X), therapy modification (D), and therapy monitoring (C). Variables recorded included patient count, admitting unit, total drugs analysed, mean drugs per patient, and mean DDIs per patient. Statistical analysis used Stata 16.0 with Student’s t-test, Z-test, and Fisher’s exact test (p&lt;0.05). <h3>Results</h3> Thirty patients (15 ICU/15 PACU) and 293 drugs were analysed (mean 9.8 drugs/patient, 95% CI: 8.9–10.7). AI identified a mean of 7.3 (6.3–8.3) interactions per patient, UpToDate 5.5 (4.2–6.9), and Medscape 5.6 (3.9–7.2); differences were not statistically significant. AI detected 218 interactions: 92.2% type C, 7.3% type D, and 0.5% type X. UpToDate identified 166 interactions: 74.7% type C (20% fewer than AI, p&lt;0.05) 24.1% type D (16.8% more than AI, p&lt;0.05) and 1.2% type X (no significant difference) Medscape detected 167 interactions: 70.7% type C (20% fewer, p&lt;0.05) 24.6% type D (17.3% more, p&lt;0.05) 4.8% type X (4.3% more, p&lt;0.05) <h3>Conclusion and Relevance</h3> No significant difference in total DDIs detected between methods was observed. However, AI overrepresented type C interactions and underreported clinically impactful interactions (types D and X). The low frequency of type X interactions warrants larger studies. AI’s variability and lack of reproducibility limit widespread clinical application. While AI shows promise as a DDI detection tool, understanding its limitations and validating its outputs remain crucial. <h3>Conflict of Interest</h3> No conflict of interest

Ähnliche Arbeiten