Repository logo
  • English
  • 中文
  • Log In
    New user? Click here to register.Have you forgotten your password?
Repository logo
    Communities & Collections
    Research Outputs
    Fundings & Projects
    People
    Organizations
    Statistics
  • English
  • 中文
  • Log In
    New user? Click here to register.Have you forgotten your password?
  1. Home
  2. .TMU Publications / 北醫出版品(教師升等著作 / 教學實踐 / 學位論文)
  3. .博碩士學位論文
  4. 113學年度
  5. 應用預訓練語言模型預測加護病房急性腎損傷患者之腎功能預後
 
  • Details
Options

應用預訓練語言模型預測加護病房急性腎損傷患者之腎功能預後

Other Title
Applying Pre-Trained Language Models to Predict Renal Prognosis in ICU Patients with Acute Kidney Injury
Type
thesis
Date Issued
2025-07-10
Author(s)
李慈恩
Advisor
張詠淳 ;廖家德
Subjects
系所名稱:醫學院人工智慧醫療碩士在職專班
Publisher
醫學院人工智慧醫療碩士在職專班
Description
學位別:碩士
口試委員:許明輝; 陳建錦; 阮逢英; 張詠淳; 廖家德
關鍵字:急性腎臟病、急性腎臟損傷、預訓練語言模型、大語言模型
Abstract
急性腎損傷(AKI)為加護病房常見且高風險之併發症,患者出院後若未及時介入治療,易進展為急性腎臟病(AKD)或慢性腎臟病(CKD)。惟目前臨床對AKD之識別與追蹤照護不足,導致早期預警信號常被忽略。為提升對AKI患者出院前AKD分期之預測能力,本研究建立一套整合結構化與非結構化資料的多模態預測架構,並探討大型語言模型(LLMs)於資料不平衡情境中的應用潛力。本研究使用MIMIC-IV v3.1資料庫,選取符合AKI診斷之成人個案,依KDIGO準則進行AKD分期,輸入特徵涵蓋生命徵象、實驗數據與病摘文字資料。模型設計包含傳統機器學習與預訓練語言模型(如Bio_Clinical_BERT),並比較SMOTE、class weight、focal loss及GPT-4、Phi-4 Mini樣本生成等平衡策略。結果顯示:結構化資料下以LightGBM、XGBoost效能最佳(Accuracy 62%);文字模型搭配GPT-4生成樣本後,Bio_Clinical_BERT Accuracy 提升至67.91%,TF-IDF結合Phi-4 Mini亦能將AKD Stage 2之F1-score提升至80%以上。整體而言,本研究驗證LLMs結合語義生成技術可有效強化少數類別預測,並提出適用於運算資源有限場域之精準且具可擴展性的解決方案,為未來AKD風險評估與智能臨床決策提供重要參考。
URI
https://203.71.86.71/handle/123456789/9357

Copyright Notice

● The digital content on this platform is part of the Taipei Medical University Institutional Repository, featuring various academic works and outputs from the institution. It offers free access to academic research and public education for non-commercial use.

● Please use the content appropriately and within legal boundaries to respect copyright owners' rights. For commercial use, please obtain prior authorization from the copyright owner. Users must not use TMUIR for any illegal purposes.

● By utilising the platform, users are deemed to have fully accepted and understood all the regulations set out in this statement, relevant laws of the Republic of China, all international internet regulations, and usage conventions.

● TMUIR is committed to protecting the interests of copyright owners. If you believe that any material on this website infringes copyright, please contact our staff at libirtmu@gmail.com, and we will remove the work from the repository.

Built with DSpace-CRIS software - Extension maintained and optimized by 4Science

  • Cookie settings
  • Privacy policy
  • End User Agreement
  • Send Feedback