基于深度神经网络的重症监护室脓毒症患者死亡风险预测模型分析

Analysis of the prediction model based on deep neural networks for mortality risk prediction for sepsis patients in intensive care units

摘要

提出基于变量选择网络(variable selection networks, VSN)和门控残差网络(gated residual networks, GRN)相结合的深度神经网络(deep neural network, DNN)模型,用于预测重症监护室(intensive care unit, ICU)脓毒症患者30 d内的死亡风险,并对模型的可解释性进行深入分析。在重症医学数据库中利用随机森林算法筛选43个重要特征,利用本文提出的模型评估死亡风险,并采用移除再训练(remove and retrain, ROAR)方法选出一种最佳的可解释性方法对结果进行解释。测试结果显示,本文提出的模型的预测性能优于其他机器学习模型,受试者工作特征曲线面积(area under receiver operating characteristic curve, AUROC)为0.967。利用ROAR方法分析中,相关性分数逐层传播(layer-wise relevance propagation, LRP)方法的AUROC从0.967下降到0.828。利用LRP方法对本文提出的模型进行可解释性分析后,确定查尔森合并症评分为最重要的特征,同时器官衰竭评分、年龄、呼吸频率也对重症监护室脓毒症患者的死亡风险有较大影响。

Abstract

A deep neural network (DNN) model is proposed by integrating variable selection networks (VSN) and gated residual networks (GRN) to predict the 30-day mortality risk of sepsis patients in the intensive care unit (ICU) and to conduct an in-depth interpretability analysis. In the critical care medical database, 43 significant features are selected using a random forest algorithm, and the proposed model is employed to evaluate mortality risk. The remove and retrain (ROAR) method is utilized to determine the optimal interpretability approach for explaining the results. Testing outcomes indicate that the proposed model outperforms other machine learning models, achieving an area under the receiver operating characteristic curve (AUROC) of 0.967. In the ROAR analysis, the AUROC of the layer-wise relevance propagation (LRP) method decreases from 0.967 to 0.828. Through interpretability analysis of the proposed model using LRP, the Charlson comorbidity score is identified as the most critical feature. In contrast, the organ failure score, age, and respiratory rate also have a pronounced impact on the mortality risk of ICU sepsis patients.

关键词

脓毒症 / 死亡风险预测 / 深度神经网络 / 移除再训练 / 相关性分数逐层传播

Key words

sepsis / mortality risk prediction / deep neural network / remove and retrain / layer-wise relevance propagation

参考文献

[1] 邱海波, 杜斌, 刘大为, 等. 全身炎症反应综合征与多器官功能障碍综合征的临床研究[J]. 中华外科杂志, 1997, 35(7): 402-405.

[2] QIU Haibo, DU Bin, LIU Dawei, et al. Clinical study on systemic inflammatory response syndrome and multiple organ dysfunction syndrome[J]. Chinese Journal of Surgery, 1997, 35(7): 402-405.

[3] GAUER R, FORBES D, BOYER N. Sepsis: diagnosis and management[J]. American Family Physician, 2020, 101(7): 409-418.

[4] WANG Yuhe, SHAN Gao, LEI Hong, et al. Prognostic impact of blood urea nitrogen to albumin ratio on patients with sepsis: a retrospective cohort study[J]. Scientific Reports, 2023, 13(1): 10013.

[5] COX D R. Regression models and life—tables[J]. Journal of the Royal Statistical Society (Series B: Methodological), 1972, 34(2): 187-220.

[6] DIAS A, GOMEZ V C, VIOLA L R, et al. Fever is associated with earlier antibiotic onset and reduced mortality in patients with sepsis admitted to the ICU[J]. Scientific Reports, 2021, 11(1): 23949.

[7] HU Jing, LÜ Chenwei, HU Xinxing, et al. Effect of hypoproteinemia on the mortality of sepsis patients in the ICU: a retrospective cohort study[J]. Scientific Reports, 2021, 11(1): 24379.

[8] VINCENT J L, MENDONCA A, CANTRAINE F, et al. Use of the SOFA score to assess the incidence of organ dysfunction/failure in intensive care units: results of a multicenter, prospective study[J]. Critical Care Medicine, 1998, 26(11): 1793-1800.

[9] FORD D W, GOODWIN A J, SIMPSON A N, et al. A severe sepsis mortality prediction model and score for use with administrative data[J]. Critical Care Medicine, 2016, 44(2): 319-327.

[10] CARRARA M, BASELLA G, FERRARIO M. Mortality prediction model of septic shock patients based on routinely recorded data[J]. Computational and Mathematical Methods in Medicine, 2015, 2015(1): 761435.

[11] KHWANNIMIT B, BHURAVANONTACHAIR R, VATTANVANIT V. Validation of the sepsis severity score compared with updated severity scores in predicting hospital mortality in sepsis patients[J]. Shock, 2017, 47(6): 720-725.

[12] HOU Nianzong, LI Mingzhe, HE Lu, et al. Predicting 30—days mortality for MIMIC—III patients with sepsis—3: a machine learning approach using XGboost[J]. Journal of Translational Medicine, 2020, 18(1): 462— 462.

[13] PERNG J W, KAO I H, KUNG C T, et al. Mortality prediction of septic patients in the emergency department based on machine learning[J]. Journal of Clinical Medicine, 2019, 8(11): 1906.

[14] 化盈盈, 张岱墀, 葛仕明. 深度学习模型可解释性的研究进展[J]. 信息安全学报, 2020, 5(3): 1-12.

[15] HUA Yingying, ZHANG Daichi, GE Shiming. Research progress in the interpretability of deep learning models[J]. Journal of Information Security, 2020, 5(3): 1-12.

[16] 陈珂锐, 孟小峰. 机器学习的可解释性[J]. 计算机研究与发展, 2020, 57(9): 1971-1986.

[17] CHEN Kerui, MENG Xiaofeng. Interpretability of machine learning[J]. Journal of Computer Research and Development, 2020, 57(9): 1971-1986.

[18] 曾春艳, 严康, 王志锋, 等. 深度学习模型可解释性研究综述[J]. 计算机工程与应用, 2021, 57(8): 1-9.

[19] ZENG Chunyan, YAN Kang, WANG Zhifeng, et al. A review of research on the interpretability of deep learning models[J]. Computer Engineering and Applications, 2021, 57(8): 1-9.

[20] RAJKOMAR A, DEAN J, KOHANE I. Machine learning in medicine[J]. New England Journal of Medicine, 2019, 380(14): 1347-1358.

[21] LIM B, ARIK S O, LOEFF N, et al. Temporal fusion transformers for interpretable multi—horizon time series forecasting[J]. International Journal of Forecasting, 2021, 37(4): 1748-1764.

[22] HOOKER S, ERHAN D, KINDERMANS P J, et al. A benchmark for interpretability methods in deep neural networks[C]∥ Advances in Neural Information Processing Systems. Vancouver: Curran Associates Incorporated, 2019.

[23] MENG C, TRINH L, XU N, et al. Interpretability and fairness evaluation of deep learning models on MIMIC—IV dataset[J]. Scientific Reports, 2022, 12: 7166.

[24] HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]// 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vega: Curran Associates Incorporated, 2016: 770-778.

[25] MITCHELL R, COOPER J, FRANK E, et al. Sampling permutations for Shapley value estimation[J]. Journal of Machine Learning Research, 2022, 23(43): 1-46.

[26] HE Shengfeng, ZHE Huang, LIU Wenxi, et al. SuperCNN: a superpixelwise convolutional neural network for salient object detection[J]. International Journal of Computer Vision, 2015, 115(3): 330-344.

[27] ZAFEIRIOU A, KALLIPOLITIS A, MAGLOGIANNIS I. Ensembling to leverage the interpretability of medical image analysis systems[J]. IEEE Access, 2023, 11: 76437-76447.

[28] SUNDARARAJAN M, TALY A, YAN Q. Axiomatic attribution for deep networks[C]// International Conference on Machine Learning. Sydney: PMLR, 2017: 3319-3328.

[29] BACH S, BINDER A, MONTAVON G, et al. On pixel—wise explanations for non—linear classifier decisions by layer—wise relevance propagation[J]. Public Library of Science One, 2015, 10(7): e0130140.

[30] HOCHULI J, HELBLING A, SKAIST T, et al. Visualizing convolutional neural network protein—ligand scoring[J]. Journal of Molecular Graphics ＆ Modelling, 2018, 84: 96-108.

[31] TOLLES J, MEURER W J. Logistic regression: relating patient characteristics to outcomes[J]. The Journal of the American Medical Association, 2016, 316(5): 533-534.

[32] OSMAN A I A, AHMED A N, CHOW M F, et al. Extreme gradient boosting (Xgboost) model to predict the groundwater levels in selangor malaysia[J]. Ain Shams Engineering Journal, 2021, 12(2): 1545-1556.

[33] MASCARO J, ASNER G P, KNAPP D E, et al. A tale of two “forests”: random forest machine learning AIDS tropical forest carbon mapping[J]. Public Library of Science One, 2014, 9(1): E85993.

[34] FRIEDMAN J H. Greedy function approximation: a gradient boosting machine[J]. The Annals of statistics, 2001, 29(5): 1189-1232.

基金资助

国家自然科学基金资助项目(61902046)

国家自然科学基金资助项目(61901074)

国家自然科学基金资助项目(62076044)

中国博士后科学基金资助项目(2021M693771)

重庆市自然科学基金资助项目(CSTB2022NSCQ-MSX0145)