- 學位論文
運用Transformer缺失數據填補技術結合時間序列和非時間序列數據進行敗血症早期預測
Using Transformer-Based Imputation Techniques to Combine Time-Series and Non-Time-Series Data for Early Prediction of Sepsis
摘要
敗血症是一種危及生命的病症，由於其復雜多變的症狀，早期預測具有挑戰性。本研究通過使用基於自注意力(Self Attention)為基礎的深度學習數值插補技術，整合不同類型的數據，提高了敗血症的預測能力。我提出了一種新的預測框架，利用深度學習進行數據插補和分類並對各種插補模型和分類網絡進行了研究，結果顯示自注意力為基礎的模型表現超越了其他模型。在早期敗血症預測方面，我的方法超越了先前的模型，能夠更有效提前七小時預測敗血症的發生。在MIMIC-IV資料庫中使用六小時的資料結合自注意力的插補和分類模型有最好的效果，在提早五小時和七小時前預測敗血症發生分別獲得AUROC 0.79 和0.83。本研究提供了一套可行性的框架，並深入比較不同插補模型的特點, 展示了自注意力為基礎的模型在資料處理和敗血症早期預測的潛力。
並列摘要
Sepsis is a life-threatening condition, and its early prediction poses challenges due to its complex and variable symptoms. In this study, I improved the predictive ability of sepsis by integrating different types of data using a deep learning numerical imputation technique based on self-attention. I proposed a novel prediction framework that utilizes deep learning for data imputation and classification. Various imputation models and classification networks were investigated, and the results showed that the self-attention-based framework outperformed other models. In terms of early sepsis prediction, my approach surpassed previous models by being able to predict the occurrence of sepsis seven hours in advance more effectively. The combination of six hours of data from the MIMIC-IV database with self-attention-based imputation and classification models yielded the best performance, achieving AUROC of 0.83 and 0.79 for five hours and seven hours early prediction of sepsis respectively. This study presents a feasible framework and provides an in-depth comparison of different imputation models, showcasing the potential of self-attention-based models in data processing and early sepsis prediction.