多模態生物標記之ARDS預測系統;A Multimodal Biomarker-Integrated Prediction System for Acute Respiratory Distress Syndrome

NCUIR > college of Health Sciences and Technology > Institute of Biomedical Engineering > Electronic Thesis & Dissertation > Item 987654321/97517

Please use this identifier to cite or link to this item: https://ir.lib.ncu.edu.tw/handle/987654321/97517

Title:	多模態生物標記之ARDS預測系統;A Multimodal Biomarker-Integrated Prediction System for Acute Respiratory Distress Syndrome
Authors:	李毓甄;LI, YU-ZHEN
Contributors:	生醫科學與工程學系
Keywords:	急性呼吸窘迫;多模態資料;機器學習;深度學習
Date:	2025-07-23
Issue Date:	2025-10-17 11:28:25 (UTC+8)
Publisher:	國立中央大學
Abstract:	急性呼吸窘迫症候群(ARDS)是一種常見且嚴重的加護病房疾病，具有高死亡率與長期功能障礙等後遺症，對病患的生存與生活品質造成重大影響。因此，如何在病程初期即早期辨識高風險患者，是臨床上非常重要的挑戰。本研究與台北醫學大學及萬芳醫院合作，蒐集170位ARDS患者的資料，包括插管初期的胸部X光影像、臨床生理數據，以及部分病患的免疫與蛋白質體資料。透過人工智慧技術，建構一套能夠整合影像與表格型資料的多模態預測系統，用以預測病患的預後結果，並進一步找出與死亡風險相關的重要臨床特徵。在模型應用方面，本研究分別導入多種影像分類模型與表格型深度學習模型進行比較，結果顯示 DenseNet121、RegNetY、DeiT 等影像模型，以及 TabularModel 與 Transformer 類型的表格模型，整體表現相對穩定且具預測潛力。研究同時開發一套操作介面，使用者可透過前端平台上傳資料並即時獲得預測結果與視覺化圖表，提升系統應用的便利性與臨床實用性。本研究成果可作為智慧醫療決策輔助系統的雛型，有助於加強臨床團隊對 ARDS 病患的風險判斷與個別化照護策略的制定。;Acute Respiratory Distress Syndrome (ARDS) is a common and severe condition frequently encountered in intensive care units (ICUs), characterized by high mortality rates and long-term functional impairments, which substantially affect patients′ survival and quality of life. Therefore, early identification of high-risk patients during the initial stages of disease progression remains a critical challenge in clinical practice. In collaboration with Taipei Medical University and Wan Fang Hospital, this study collected data from 170 ARDS patients, including chest X-ray images obtained during the early intubation period, clinical physiological parameters, as well as immunological and proteomic data for a subset of the patients. Leveraging artificial intelligence techniques, a multimodal prediction system was developed to integrate imaging and tabular data for prognostic prediction, while simultaneously identifying key clinical features associated with mortality risk. In terms of model application, various imaging classification models and tabular deep learning models were implemented and compared. Results demonstrated that models such as DenseNet121, RegNetY, and DeiT for imaging data, and TabularModel and Transformer-based models for tabular data, exhibited relatively stable performance with promising predictive capabilities. Additionally, a user-friendly interface was developed, enabling users to upload data through a front-end platform and receive real-time predictions and visualized outputs, thereby enhancing system usability and clinical applicability. The outcomes of this study may serve as a prototype for intelligent clinical decision support systems (CDSS), facilitating improved risk assessment and personalized care strategies for ARDS patients by clinical teams.
Appears in Collections:	[Institute of Biomedical Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	11	View/Open

社群 sharing

Loading...