Emergency Management of Permanent Tooth Avulsion: Comparative Performance of Dental Students and Artificial...

To pioneer the first comprehensive benchmarking of dental students against 22 multimodal Artificial Intelligence (AI) models spanning six major architectures (Anthropic, DeepSeek, Google Gemini, OpenAI, Meta Llama, Mistral/Qwen) in emergency permanent tooth avulsion management, quantifying performance differences across Bloom's Taxonomy higher-order cognitive domains (Apply, Analyse, Evaluate) against International Association of Dental Traumatology (IADT) guidelines to optimise evidence-based educational integration.

This cross-sectional study compared 35 fifth-year dental students with 22 multimodal AI models grouped by producer (Anthropic, DeepSeek, Google Gemini, OpenAI, Meta, Mistral/Qwen) using a validated clinical vignette featuring a 14-year-old with three avulsed mature permanent teeth (#11, 12, 22) with clinical photograph. Performance was assessed through four questions (1 Apply, 2 Analyse, 1 Evaluate) via an IADT-aligned rubric validated against paediatric dentistry standards. Responses were independently scored by two blinded evaluators. Subsequently, scores were compared using ANOVA with post hoc Tukey HSD, Kruskal-Wallis, and chi-square tests (SPSS v20, α = 0.05).

Overall performance analysis revealed significant group differences (p = 0.012, partial η² = 0.27). Only students achieved high-acceptability A-level performance (17.1%, χ² = 39.27, p = 0.003), despite top AI models (Claude-Sonnet-3.7-Reasoning, DeepSeek-R1: 7.0/10) marginally exceeding student mean scores (6.5/10). Performance varied among AI systems, with students significantly outperforming Meta Llama models (mean difference = 3.50, p = 0.014), which showed exclusively not acceptable (D-level) performance. Students dominated Bloom's Analyse level (visual diagnosis: 100% tooth identification vs. 77.3% AI failure; age-specific management differentiation, p < 0.001), while AI excelled in Apply/Emergency steps (medical management: 81.8% perfect scores) and Evaluate/prosthetic contraindication reasoning (Anthropic: 0.92 ± 0.20 vs. students: 0.41 ± 0.33, p = 0.005). Critical AI deficiencies included tooth misidentification (77% of models), failure to mention follow-up schedules for immature replanted teeth (100%), and inappropriate intervention recommendations (40.9% suggested unsuitable options like resin-bonded bridges). For soft tissue management, only 23% of AI models addressed the need to suture gingival lacerations.

AI's fundamental limitations in visual diagnosis, IADT misinterpretations, and inadequate clinical translation mandate strict positioning as a supplemental protocol accelerator-not an independent diagnostic tool. Educational integration requires: (1) Careful model selection and supervision for any visual triage applications; (2) Mandatory IADT verification; and (3) Preservation of human clinical reasoning as the irreplaceable cornerstone of permanent tooth avulsion management.