Table 2.

Prescription-review performance of unaided pharmacist review vs human-AI collaborative review on 213 outpatient prescriptions sampled at a tertiary teaching hospital in China (October 2025-November 2025) in a 2-period crossover study, using an independent expert reference standard (62 inappropriate prescriptions)^a.

Groups	Accuracy (%)	Sensitivity (%)	Specificity (%)	False-positive rate (%)	False-negative rate (%)
Pharmacist alone	82.6	55	94	6	45
Human-AI collaborative	97.2	98	96.7	3.3	2
Chi-square test (df)	25.71 (1)	25.04 (1)	—b	—	25.04 (1)
P value	<.001	<.001	.29	.29	<.001

^aComparisons used paired McNemar tests with Yates continuity correction (213 paired observations). For accuracy, there were 33 discordant pairs that were correct only with collaboration and 2 that were correct only when unaided; for sensitivity and the false-negative rate, there were 27 and 0 discordant pairs, respectively. Accuracy denominators were 213; sensitivity, specificity, and the error rates used 62 inappropriate and 151 appropriate prescriptions, respectively.

^bSpecificity and false-positive-rate differences were not significant, and their test statistics are not reported.