Table 3.
Impact of retrieval-augmented generation (RAG) integration on diagnostic accuracy across different clinical personas (n=42).
| Persona and metrics | Base model, n (%, 95% CI) | RAG model, n (%, 95% CI) | P value (McNemar test) | |||||
| Radiologist | ||||||||
|
|
Top-1 accuracy | 12 (29%, 95% CI 16%-45%) | 6 (14%, 95% CI 5%-29%) | .03a | ||||
|
|
Top-3 accuracy | 15 (36%, 95% CI 22%-52%) | 10 (24%, 95% CI 12%-39%) | .13 | ||||
| Ortho oncologist | ||||||||
|
|
Top-1 accuracy | 8 (19%, 95% CI 9%-34%) | 7 (17%, 95% CI 7%-31%) | .65 | ||||
|
|
Top-3 accuracy | 14 (33%, 95% CI 20%-50%) | 14 (33%, 95% CI 20%-50%) | >.99 | ||||
aStatistically significant (P<.05). P values indicate the difference in diagnostic accuracy before and after RAG integration within the same persona, calculated using the standard McNemar test. The baseline top-3 accuracy between the base-radiologist and base-ortho oncologist showed no significant difference (P=.76).