Large Language Models for NiTi File-Group Selection in Endodontics: A Two-Stage Comparative Study

This study evaluated whether a structured nickel-titanium file-group framework improves large language model recommendations for endodontic instrumentation. Five models were tested across 48 standardised scenarios under label-only and structured-framework conditions. Two independent endodontists, blinded to model identity, explicit stage labels and each other's ratings, scored 480 responses. The Clinical Appropriateness Score, a study-specific expert-rated measure on a 0-100 scale, increased from 75.0 to 80.6 after framework provision. Strict concordance with the expert-preferred group rose from 24.6% to 37.9%, but this gain was concentrated among responses containing prohibited proprietary names, which increased from 14.2% to 40.8%. Overall, 5.2% of responses were non-acceptable and the reduction from 7.1% to 3.3% was not statistically significant. Structured prompting improved recommendation quality, but did not produce a detectable concordance gain among terminology-compliant responses. Expert oversight remains essential.