Scholarlink

Temporomandibular degenerative joint disease (TM DJD) is a multifactorial condition with complex clinical presentations. This study presents a multimodal framework centered on structured summarization of clinical text, supported by imaging information from automatically registered MRI and CBCT scans. Two large language models, BART and DeepSeek-R1, were fine-tuned on 1,813 annotated text segments from 500 TM DJD patient records to extract 56 clinical indicators, including pain severity, jaw function, imaging findings, and sleep disturbances. The models converted narrative notes into structured data fields for use in clinical dashboards enabling patient-specific and population-level analyses. BART outperformed DeepSeek in clinical field extraction accuracy, precision, and recall, despite DeepSeek achieving slightly higher ROUGE metrics based on word-level overlap. A parallel automated MRI-to-CBCT registration pipeline achieved submillimeter accuracy and a 98.75% success rate. This work extracted clinically meaningful pain comorbidities and radiological findings from unstructured clinical narratives, enabling actionable insights for musculoskeletal precision care. The future integration of structured clinical data and multimodal image analyses may enable holistic, personalized patient models.

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

AI-Driven Multimodal TMJ Patient Modeling: From Unstructured Notes to Precision Treatment

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Alban Gaydamour,
  • Enzo Tulissi,
  • Claudia Mattos,
  • Rodrigo Teixeira,
  • Maxwell Shin,
  • Adam Hershey,
  • Anabelle Kwon,
  • Felicia Miranda,
  • Marcela Gurgel,
  • Selene Barone,
  • Aron Aliaga,
  • Marilia Yatabe,
  • Paulo Zupelari,
  • Marina Zupelari,
  • David Hanauer,
  • Nina Hsu,
  • Steve Pieper,
  • Eduardo Caleme,
  • Jonas Bianchi,
  • Joao Goncalves,
  • Daniela Goncalves,
  • Lawrence Wolford,
  • Antonio Ruellas,
  • Juan Prieto,
  • Tengfei Li,
  • Hongtu Zhu,
  • Runpeng Dai,
  • Martin Styner,
  • Najla Al Turkestani,
  • Alexandre F. DaSilva,
  • Lucia Cevidanes

摘要

Temporomandibular degenerative joint disease (TM DJD) is a multifactorial condition with complex clinical presentations. This study presents a multimodal framework centered on structured summarization of clinical text, supported by imaging information from automatically registered MRI and CBCT scans. Two large language models, BART and DeepSeek-R1, were fine-tuned on 1,813 annotated text segments from 500 TM DJD patient records to extract 56 clinical indicators, including pain severity, jaw function, imaging findings, and sleep disturbances. The models converted narrative notes into structured data fields for use in clinical dashboards enabling patient-specific and population-level analyses. BART outperformed DeepSeek in clinical field extraction accuracy, precision, and recall, despite DeepSeek achieving slightly higher ROUGE metrics based on word-level overlap. A parallel automated MRI-to-CBCT registration pipeline achieved submillimeter accuracy and a 98.75% success rate. This work extracted clinically meaningful pain comorbidities and radiological findings from unstructured clinical narratives, enabling actionable insights for musculoskeletal precision care. The future integration of structured clinical data and multimodal image analyses may enable holistic, personalized patient models.