Scholarlink

Background <p>This study evaluated the performance of a multimodal large language model (MLLM), Chat Generative Pretrained Transformer-5.5 (ChatGPT-5.5), in determining orthodontic treatment need based on the Index of Complexity, Outcome, and Need (ICON) using intraoral photographs and digital model images. Agreement at the ICON component level and consistency with the total ICON score were also examined.</p> Methods <p>A total of 104 patients contributed 520 intraoral photographs and 520 rendered digital model images. Two orthodontists independently scored all ICON components (aesthetics, crowding/spacing, crossbite, vertical relationship, and buccal relationship), and cases with complete interrater agreement were used as the reference standard. A standardized ICON prompt for ChatGPT-5.5 was optimized using the Iterative Prompt Calibration (IPC) method. Model outputs were assessed using accuracy, sensitivity, specificity, F1-score, exact agreement rates, Cohen’s kappa, and intraclass correlation coefficients (ICC). Confusion matrices and ROC curves were generated for the binary treatment-need decision.</p> Results <p>The ChatGPT-5.5 achieved overall accuracies of 74.0% for intraoral photographs and 72.1% for model images. Recall was high for treatment-required cases (0.871 and 1.00, respectively) but markedly lower for cases not requiring treatment (0.471 and 0.147). Component-level agreement varied considerably, with significant differences between photographs and models for crowding/spacing and vertical relationships. The anteroposterior relationship showed higher accuracy in model evaluations. ICC values for the total ICON score were 0.463 for photographs (moderate reliability) and 0.154 for models (poor reliability). AUC values were 0.671 and 0.574, indicating limited discriminative performance.</p> Conclusions <p>Although ChatGPT-5.5, evaluated as a MLLM, demonstrated moderate accuracy in the binary classification of orthodontic treatment need, its inconsistent performance across ICON components and tendency to overestimate treatment need limit its current clinical reliability. Therefore, clinical decision-making should remain guided by expert judgment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Diagnostic performance of a multimodal large language model in assessing the index of complexity, outcome, and need: comparison with orthodontist evaluation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Rumeysa Bilici Geçer,
  • Arda Tabancalı,
  • Çağla Hasgül,
  • Elif İpek Komar,
  • Emine Selen Sarıoğlu,
  • Derya Dursun

摘要

Background

This study evaluated the performance of a multimodal large language model (MLLM), Chat Generative Pretrained Transformer-5.5 (ChatGPT-5.5), in determining orthodontic treatment need based on the Index of Complexity, Outcome, and Need (ICON) using intraoral photographs and digital model images. Agreement at the ICON component level and consistency with the total ICON score were also examined.

Methods

A total of 104 patients contributed 520 intraoral photographs and 520 rendered digital model images. Two orthodontists independently scored all ICON components (aesthetics, crowding/spacing, crossbite, vertical relationship, and buccal relationship), and cases with complete interrater agreement were used as the reference standard. A standardized ICON prompt for ChatGPT-5.5 was optimized using the Iterative Prompt Calibration (IPC) method. Model outputs were assessed using accuracy, sensitivity, specificity, F1-score, exact agreement rates, Cohen’s kappa, and intraclass correlation coefficients (ICC). Confusion matrices and ROC curves were generated for the binary treatment-need decision.

Results

The ChatGPT-5.5 achieved overall accuracies of 74.0% for intraoral photographs and 72.1% for model images. Recall was high for treatment-required cases (0.871 and 1.00, respectively) but markedly lower for cases not requiring treatment (0.471 and 0.147). Component-level agreement varied considerably, with significant differences between photographs and models for crowding/spacing and vertical relationships. The anteroposterior relationship showed higher accuracy in model evaluations. ICC values for the total ICON score were 0.463 for photographs (moderate reliability) and 0.154 for models (poor reliability). AUC values were 0.671 and 0.574, indicating limited discriminative performance.

Conclusions

Although ChatGPT-5.5, evaluated as a MLLM, demonstrated moderate accuracy in the binary classification of orthodontic treatment need, its inconsistent performance across ICON components and tendency to overestimate treatment need limit its current clinical reliability. Therefore, clinical decision-making should remain guided by expert judgment.