Multimodal artificial intelligence in dental diagnosis: influence of reasoning depth and domain-specific medical...

Large language models are increasingly explored as clinical decision-support tools, but the relationship between configurable reasoning effort and multimodal diagnostic performance remains unclear. This study evaluated the association between API-configured reasoning effort and multimodal oral and dental diagnostic performance in GPT-5 and compared its outputs with those of the compact open-weight medical-domain model MedGemma-1.5-4b-it. Eight anonymized multimodal case reports representing dental trauma, odontogenic cysts, odontogenic tumors, implant-associated infection, oral mucosal disorders, hereditary mucosal disease, and restorative management were reformatted into structured evaluation scenarios. Final diagnostic labels, definitive treatment plans, author interpretations, answer-revealing captions, and postoperative outcomes were excluded to reduce answer leakage. Five configurations were evaluated: GPT-5 under four API-configured reasoning-effort settings (high, medium, low, and minimal) and MedGemma-1.5-4b-it as a compact medical-domain comparator. Outputs were assessed using a six-domain 0-5 rubric covering diagnostic accuracy, diagnostic completeness, reasoning quality, treatment planning, safety and risk awareness, and scope adherence. Diagnostic scores were checked against the exact archived outputs and source reference diagnoses before final analysis, and a domain-level Safety Override was applied to clinically dangerous recommendations. Repeated-measures differences were analyzed at the case level, with complementary exploratory analyses of primary-diagnosis agreement, reference-diagnosis inclusion, clinical-pathway appropriateness, and safety. Inter-rater reliability of the original independent blinded ratings was excellent (intraclass correlation coefficient = 0.92; 95% confidence interval, 0.87-0.95). Total scores differed across configurations (Friedman χ²(4) = 27.24, p = 0.000018; Kendall's W = 0.85). Mean scores were 28.0 ± 2.5 for GPT-5 high, 27.8 ± 1.6 for medium, 22.8 ± 4.0 for low, 18.9 ± 2.8 for minimal, and 17.1 ± 3.3 for MedGemma-1.5-4b-it. Primary-diagnosis agreement with the published reference diagnosis occurred in 6/8 cases for GPT-5 high, 6/8 for medium, 5/8 for low, 4/8 for minimal, and 1/8 for MedGemma; reference-diagnosis inclusion either as the primary diagnosis or within the differential occurred in 8/8, 8/8, 8/8, 7/8, and 3/8 cases, respectively. One safety-critical error occurred in GPT-5 low; none occurred in the other configurations. GPT-5 high and medium had the highest observed mean scores, but no pairwise Wilcoxon comparison remained significant after Holm correction. Primary-diagnosis agreement, reference-diagnosis inclusion, clinical-pathway appropriateness, and safety provided complementary views of open-ended clinical performance. MedGemma showed lower primary-diagnosis agreement and reference-diagnosis inclusion, while no MedGemma response met the final source-verified safety-critical criterion. The between-family comparison remains a pragmatic comparison of distinct deployment paradigms rather than a controlled test of training-domain specialization. In this dataset, higher GPT-5 reasoning-effort settings were associated with higher observed multimodal performance, particularly in diagnostically complex cases. Reasoning-effort configuration may therefore be an important design consideration for future dental decision-support systems, but the association is descriptive and requires confirmation in larger prospective studies.