利用機器學習方法於 MIMIC 資料庫之早期敗血症預測

Early Prediction of Sepsis Using Machine Learning Methods on MIMIC Database

吳承彥(Cheng-Yan Wu)

指導教授 ： 周呈霙

國立臺灣大學/共同教育中心/統計碩士學位學程/碩士(2022年)

https://doi.org/10.6342/NTU202202105

摘要

敗血症是加護病房內嚴重疾病之一，此疾病發病後可能導致患者的高死亡率和多種併發症。由於不同的患者、生命特徵、敗血症標準和預測方法，敗血症的早期預測是具有挑戰性。本研究旨在通過機器學習算法和深度學習方法開發一種高準確率的早期敗血症預測模型，該模型可以提高敗血症的早期預測，藉此警示醫生那些未來可能發展成敗血症的病患，從而降低發病率和死亡率。
此研究所開發的模型分類結果顯示 XGBoost 和 CNN 預測模型在分類敗血症方面表現出很強的性能。在 MIMIC-III 資料庫中，使用 SIRS 標準和XGBoost模型在敗血症發病時的病患 AUROC 約為 0.876，發病前 8 小時的 AUROC 為 0.780。使用 qSOFA標準在敗血症發病時的病患 AUROC 約為 0.942，發病前 8 小時的 AUROC 為0.729。CNN 預測模型使用 SIRS 標準在敗血症發病時達到了 0.996 AUROC，在發病前 8 小時的 AUROC 值為 0.945。
在 MIMIC-IV 資料庫中，使用 SIRS 標準和XGBoost模型在敗血症發病時的病患 AUROC 約為0.836，發病前 8 小時的 AUROC 為 0.902。使用 qSOFA 標準在敗血症發病時的病患 AUROC 約為 0.823，發病前 8 小時的AUROC 為 0.737。 CNN 預測模型使用 SIRS 標準在敗血症發病時達到了 0.992 的AUROC，在發病前 8 小時的 AUROC 值為 0.917.
和前人做法不同的地方是我將一般的數值輸入轉換成圖像輸入，並且使用圖像輸入比起數值輸入可以得到更好的分類效果。因此，相信 CNN 和 XGBoost 預測模型可以用於提前 8小時預測敗血症發作。根據本研究的結果，CNN 和 XGBoost 預測模型可以使用八個特徵提前8小時準確預測敗血症發作。此外，僅使用八個特徵就獲得了這些高準確率的早期預測結果。總之，結果顯示 CNN 和 XGBoost 預測模型在敗血症的早期預測上可以得到很好的效果。

關鍵字

敗血症 ； 早期預測 ； 加護病房 ； 機器學習 ； 深度學習

並列摘要

Sepsis is one of the severe diseases which has high mortality, multiple complications, and cost overruns among patients treated in the intensive care unit (ICU). Because of variations in different patient cohorts, clinical variables, sepsis criterion, and prediction tasks, early clinical recognition of sepsis is challenging. This study aimed to develop a high-performance early sepsis prediction model by a machine learning algorithm and deep learning method that can improve the early detection of sepsis, thereby reducing morbidity and mortality.
The model classification results developed in this study show that the XGBoost and CNN prediction models exhibit strong performance in classifying sepsis. In the MIMIC-III dataset, subjects using the SIRS criterion and the XGBoost model had an AUROC of approximately 0.876 at the onset of sepsis and an AUROC of 0.780 eight hours before onset. Using the qSOFA criterion had an AUROC of 0.942 at the onset of sepsis and an AUROC of 0.729 eight hours before onset. The CNN prediction model achieved 0.996 AUROC at the onset of sepsis using the SIRS criterion and an AUROC value of 0.945 eight hours before the onset of sepsis.
In the MIMIC-IV dataset, using the SIRS criterion and the XGBoost model had an AUROC of 0.836 at the onset of sepsis and an AUROC of 0.902 eight hours before onset. Subjects using the qSOFA criterion had an AUROC of approximately 0.823 at the onset of sepsis and an AUROC of 0.737 eight hours before onset. Using the SIRS criterion, the CNN prediction model achieved an AUROC of 0.992 at the onset of sepsis and an AUROC value of 0.917 eight hours before the onset of sepsis.
According to the results, using eight features, the CNN and XGBoost prediction models could accurately predict sepsis onset up to eight hours in advance. Our model significantly outperformed previously existing ones. Furthermore, these high-accuracy early prediction results were obtained using only eight features. In summary, results demonstrated that the CNN and XGBoost prediction models could improve early sepsis detection.

並列關鍵字

Sepsis ； Early prediction ； Intensive care unit ； Machine learning ； Deep learning

參考文獻

[1] Pierangela Bruno and Francesco Calimeri. Using heatmaps for deep learning based disease classification. In 2019 IEEE conference on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB), pages 1–7. IEEE, 2019.

[2] Paul E Marik and Abdalsamih M Taeb. Sirs, qsofa and new sepsis definition. Journal of Thoracic Disease, 9(4):943, 2017.

[3] Eamon P Raith, Andrew A Udy, Michael Bailey, Steven McGloughlin, Christopher MacIsaac, Rinaldo Bellomo, David V Pilcher, et al. Prognostic accuracy of the sofa score, sirs criteria, and qsofa score for in-hospital mortality among adults with suspected infection admitted to the intensive care unit. JAMA, 317(3):290–300, 2017.

[4] Andrew Lever and Iain Mackenzie. Sepsis: definition, epidemiology, and diagnosis. BMJ, 335(7625):879–883, 2007.

[5] Elizabeth K Stevenson, Amanda R Rubenstein, Gregory T Radin, Renda Soylemez Wiener, and Allan J Walkey. Two decades of mortality trends among patients with severe sepsis: a comparative meta-analysis. Critical Care Medicine, 42(3):625, 2014.