請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/86987
標題: | 運用自然語言處理合併結構化資料預測術後死亡率 Predicting Postoperative Mortality with Structured Data and Natural Language Processing |
作者: | 陳沛甫 Pei-Fu Chen |
指導教授: | 賴飛羆 Feipei Lai |
關鍵字: | 機器學習,深度神經網路,自然語言處理,圍手術期風險,預測模型, machine learning,deep neural network,natural language processing,perioperative risk,prediction model, |
出版年 : | 2022 |
學位: | 博士 |
摘要: | 近年來,機器學習和深度學習運用電子醫療系統的資料,有效地預測醫療中的風險發生。在臺灣,我們有每位病人的術前麻醉評估資料和整個醫療過程的電子病歷紀錄。在日新月異的機器學習和自然語言處理發展下,這些資料可用作輸入特徵訓練模型來更準確地預測風險。術前診斷和手術方式的文字描述可以幫助麻醉科醫師了解病人手術相關風險,但運用自然語言處理進行術式和診斷的文字的風險預測還有待研究。因此,本計畫預計使用患者的圍手術期的資料,使用機器學習和自然語言處理建立模型,預測病人的死亡率。
本回溯性研究收案包含所有接受全身或脊椎麻醉之受試者,取得術前資料作為輸入的特徵,包括病人基本資料、共病症、實驗室檢查資料以及診斷和手術方式的文字資料,訓練模型學習是否三十天內院內死亡。資料類型包括類別變項、連續變項和文字資料,運用機器學習(random forest, eXtreme Gradient Boost, logistic regression)、深度學習和自然語言處理進行訓練。先比較自然語言處理中Bidirectional Encoder Representations from Transformers (BERT)、Embeddings from Language Models 和 Global Vector等方法,表現最好的方法再用來與deep neural network (DNN) 結合成混成模型。先比較機器學習和DNN用一樣的輸入特徵的表現,接著計算 area under the receiver operating characteristic curve (AUROC) 和area under the precision-recall curve (AUPRC)比較加入文字資料的效果。 在語言模型的比較部分,運用術前診斷和手術名稱預測術後死亡時,bio-clinical BERT模型相較於其他語言模型有更高的AUROC (0.883)。在比較合併結構化與非結構化資料的模型部分,BERT-DNN模型有最高的AUROC (0.964) 和 AUPRC (0.336)。BERT-DNN的AUROC顯著性地高於eXtreme Gradient Boost classifier、logistic regression和American Society of Anesthesiologists physical status,但不顯著性高於DNN和random forest classifier。BERT-DNN的AUPRC顯著性地高於其他模型。 描述手術的文字對於預測術後死亡率很重要。本研究將術前診斷和手術方法運用BERT語言模型轉成向量,並且與深度神經網路結合用以預測術後死亡率。這個預測模型能夠用結構化資料和文字辨識高風險病患族群,可以減少遺漏,並且及早介入處置、溝通和管理醫療資源。 Using machine learning techniques has resulted in more accurate predictions of postoperative mortality compared to previous methods. Before a patient undergoes a surgery, we had free text descriptions of the preoperative diagnosis and the planned procedure. Since reading these descriptions can assist anesthesiologists in assessing the risk of the surgery, we hypothesized that deep learning models utilizing unstructured text can enhance postoperative mortality prediction. However, it can be difficult to extract useful concept embeddings from unstructured clinical text. The goal of this study is to develop a deep learning model that combines structured and unstructured features in order to predict the risk of 30-day postoperative mortality within a hospital before surgery. The effectiveness of machine learning models that use preoperative data, with or without free clinical text, in predicting postoperative mortality will be evaluated. In this study, we retrospectively gathered discharge summaries, surgical information, and preoperative anesthesia assessment of patients who received general or neuraxial anesthesia from electronic medical records. We first compared different natural language processing methods, including Global Vector, Embeddings from Language Models, and Bidirectional Encoder Representations from Transformers (BERT). Then, we combined the top-performing aforementioned method with a deep neural network (DNN) model to extract information from clinical texts. We compared the performance of machine learning models, including random forest, eXtreme gradient boost, and logistic regression, to the DNN model using the same input features. We assessed the impact of adding text information on model performance by measuring the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). Statistical significance was determined using P < .05. In comparing performance of embedding methods, the bio-clinical BERT model had the highest AUROC of 0.883 to predict postoperative mortality with text of preoperative diagnosis and procedures among all language models. In comparing performance of combined structured and unstructured models, BERT-DNN had the highest AUROC of 0.964 (95% confidence interval [CI]: 0.961 – 0.967) and the highest AUPRC of 0.336 (95% CI: 0.276 – 0.402). The BERT-DNN was significantly higher than the eXtreme gradient boost classifier, logistic regression, and American Society of Anesthesiologists physical status in AUROC, but not significantly higher than the DNN and random forest classifier. The BERT-DNN was significantly higher than other models in AUPRC. In summary, the surgical text descriptions were crucial for predicting postoperative mortality. In this study, we used deep learning models to combine the word embeddings of preoperative diagnoses and planned procedures, obtained using the contextualized language model BERT, to predict postoperative mortality. This ability to predict risk can help identify patients who are at higher risk based on the structured data and text in electronic health records, potentially reducing missed opportunities for early intervention, communication, and resource management. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/86987 |
DOI: | 10.6342/NTU202210204 |
全文授權: | 未授權 |
顯示於系所單位: | 生醫電子與資訊學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-1.pdf 目前未授權公開取用 | 3.71 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。