Abstract / Summary
We evaluated a secure medical generative artificial intelligence (AI) environment to verify active ingredient dose equivalence between patients’ own medications and hospital formulary alternatives during the reconciliation of patients’ own medications. Data for patients admitted between January 1 and March 31, 2026, and initially entered into a Microsoft Access-based checking system were analyzed using manual pharmacist calculations as the reference standard. For single-ingredient pairs, the Access system and an AI-alone condition were assessed; for fixed-dose combination pairs, only the AI-alone condition was assessed. Calculated values and final judgments were compared. As an exploratory analysis, an AI condition supplied with investigator-verified composition information from electronic package inserts was applied to discrepant single-ingredient pairs and all fixed-dose combination pairs. Among 6,866 single-ingredient pairs, the Access system achieved sensitivity and specificity of 100.00％ and 99.88％, respectively, while the AI-alone condition yielded a sensitivity of 98.44％ and a specificity of 99.99％. For 151 fixed-dose combination pairs, specificity was low (54.00％) because many equivalent pairs were misclassified as non-equivalent; sensitivity was not calculated because the reference standard identified only one non-equivalent pair. AI-alone discrepancies were mainly attributable to errors in extracting composition information, and repeated analyses of the 18 discrepant single-ingredient pairs reproduced the same discrepancies. When investigator-verified composition information was provided, calculated values and final judgments were concordant with manual calculations for all 18 discrepant single-ingredient pairs and all 151 fixed-dose combination pairs. These findings suggest that externally provided composition information and pharmacist review are important when using medical generative AI for this task.