An Interpretable Machine Learning Model for Predicting In-hospital Death Risk in Patients with Cardiac Arrest: Based on US Medical Information Mart for Intensive Care Database Ⅳ 2.0
-
摘要:
目的 构建可预测心脏骤停患者住院期间死亡风险的机器学习模型,并对其进行解释。 方法 提取美国重症监护医学信息数据库Ⅳ(Medical Information Mart for Intensive Care database Ⅳ,MIMIC-Ⅳ)2.0中心脏骤停患者转入ICU 24 h内首次临床资料及住院期间转归,基于机器学习算法构建6种可预测心脏骤停患者院内死亡风险的模型,包括XGBoost模型、轻量级梯度提升机(light gradient boosting machine, LGBM)模型、决策树(decision tree, DT)模型、K近邻(K-nearest neighbor,KNN)模型、Logistic回归模型、随机森林(random forest, RF)模型。采用受试者操作特征(receiver operator characteristic, ROC)曲线、临床决策曲线及校准曲线对模型进行评价,并采用Shapley加性解释(Shapley additive explanation, SHAP)算法评估不同临床特征对最优模型的影响,以增加模型的可解释性。 结果 共1465例符合纳入与排除标准的心脏骤停患者入选本研究。其中住院期间存活773例、死亡692例。经筛选,共纳入82个临床特征用于机器学习模型构建。模型评价结果显示,相较于其余5种模型,LGBM模型预测心脏骤停患者院内死亡的曲线下面积(area under the curve,AUC)更高[0.834(95% CI: 0.688~0.894)],且相对于Logistic回归模型、XGBoost模型,其对死亡风险的预测准确性更高(校准度:0.166),临床决策性能更优,整体性能最佳。SHAP算法分析显示,对LGBM模型输出结果影响最大的3个临床特征分别为格拉斯哥睁眼反应评分、碳酸氢盐水平、白细胞计数。 结论 基于大型公共医疗卫生数据库建立的可预测心脏骤停患者住院期间死亡风险的机器学习模型中,LGBM模型性能最优,其可辅助临床进行更高效的疾病管理和更精准的医疗干预。 -
关键词:
- 心脏骤停 /
- 预测模型 /
- 机器学习 /
- SHAP算法 /
- 美国重症监护医学信息数据库
Abstract:Objective To develop and validate an interpretable machine learning model based on clinical characteristics to predict the risk of in-hospital death in patients with cardiac arrest. Methods First clinical data of cardiac arrest patients admitted to ICU within 24 h and outcomes during hospitalization were extracted from Medical Information Mart for Intensive Care database Ⅳ (MIMIC-Ⅳ) 2.0. Six models predicting in-hospital death risk of cardiac arrest patients were constructed based on machine learning algorithm: XGBoost model, light gradient boosting machine (LGBM) model, decision tree (DT) model, K-nearest neighbor (KNN) model, Logistic regression model, and random forest (RF) model. Receiver operator characteristic (ROC) curve, clinical decision curve and calibration curve were used to evaluate the 6 models. Shapley additive explanation (SHAP) algorithm was used to explain and evaluate the effects of different clinical features on the optimal model to increase its interpretability. Results A total of 1465 patients with cardiac arrest who met inclusion and exclusion criteria were included in the study. Among them, 773 patients survived and 692 died during hospitalization. After screening, a total of 82 clinical features were included for machine learning model construction. Compared with the other five models, the LGBM model had a higher area under the curve for predicting in-hospital death in cardiac arrest patients [0.834(95% CI: 0.688-0.894)], higher prediction accuracy for the risk of death than the Logistic regression model and XGBoost model (calibration degree: 0.166), better clinical decision performance, and displayed optimal overall performance. SHAP algorithm analysis showed that the three clinical features that had the greatest impact on the output of LGBM model were Glasgow eyes score, bicarbonate level and white blood cell count. Conclusion Based on a large public medical and health database, a machine learning model named LGBM has the best performance to predict the risk of in-hospital death in patients with cardiac arrest, which will be helpful to assist more efficient clinical disease management and more precise medical intervention. 作者贡献:龚欢欢负责数据统计、图表绘制及论文撰写;柯晓伟负责数据整理及论文修订;李湘民、王爱民负责研究设计及写作指导。利益冲突:所有作者均声明不存在利益冲突 -
图 3 预测效能Top 3模型的校准曲线
LGBM:同图 2
图 4 预测效能Top 3模型的临床决策曲线
LGBM:同图 2
图 5 SHAP汇总图
A.不同临床特征对模型输出结果影响性的SHAP值;B.各临床特征的平均SHAP绝对值
SHAP:Shapley加法解释; GCS: APSⅢ、SOFA: 同表 1表 1 1465例心脏骤停患者基线主要临床资料
指标 死亡组(n=692) 存活组(n=773) P值 年龄(x±s,岁) 67.57±16.35 65.57±16.15 0.019 女性[n(%)] 286(41.3) 290(37.5) 0.136 心率(x±s,次/min) 93.66±22.36 87.44±22.19 <0.001 呼吸频率(x±s,次/min) 21.25±6.50 20.01±6.22 <0.001 体温[M(P25, P75), ℃] 36.50(35.80,36.89) 36.72(36.33,37.06) <0.001 糖尿病[n(%)] 113(16.3) 141(18.2) 0.335 心力衰竭[n(%)] 248(35.8) 345(44.6) 0.001 肾衰竭[n(%)] 395(57.1) 340(44.0) <0.001 肾脏替代治疗[n(%)] 97(14.0) 74(9.6) 0.008 使用多巴胺[n(%)] 125(18.1) 91(11.8) 0.001 使用肾上腺素[n(%)] 147(21.2) 113(14.6) 0.001 GCS评分[M(P25, P75), 分] 10(3,15) 13(9,15) <0.001 SOFA评分(x±s,分) 9.82±4.32 7.60±4.38 <0.001 LODS评分(x±s,分) 9.24±3.80 6.59±3.75 <0.001 APSⅢ评分(x±s,分) 81.63±29.39 59.34±27.76 <0.001 SIRS评分[M(P25, P75), 分] 3(3,4) 3(2,3) <0.001 住院时间[M(P25, P75), d] 5.76(2.77,11.28) 11.88(6.97,22.05) <0.001 GCS:格拉斯哥昏迷评分;SOFA: 序贯器官功能衰竭评价;LODS:器官功能障碍逻辑性评分;APSⅢ:急性生理学评分系统Ⅲ;SIRS:全身炎症反应综合征 表 2 6种机器学习模型预测心脏骤停患者院内住院死亡风险的性能比较
预测模型 AUC(95% CI) 灵敏度(95% CI, %) 特异度(95% CI, %) PLR(95% CI) NLR(95% CI) KNN模型 0.748(0.659~0.815) 71(67.4~82.2) 63(57.7~73.6) 1.93(1.662~2.215) 0.46(0.152~0.694) DT模型 0.687(0.581~0.778) 61(59.3~75.5) 74(66.9~83.9) 2.43(1.790~2.683) 0.52(0.263~0.821) RF模型 0.776(0.650~0.820) 59(55.1~72.9) 84(66.8~88.4) 3.77(2.825~4.556) 0.48(0.189~0.701) Logistic回归模型 0.809(0.661~0.853) 76(66.8~84.7) 75(52.2~81.4) 3.06(2.549~4.018) 0.32(0.125~0.682) XGBoost模型 0.827(0.679~0.875) 75(64.5~80.1) 78(63.8~82.0) 3.36(2.699~4.343) 0.32(0.119~0.668) LGBM模型 0.834(0.688~0.894) 70(63.2~81.0) 81(65.9~85.4) 3.64(2.776~4.626) 0.37(0.165~0.694) KNN、DT、RF、LGBM:同图 2;AUC:曲线下面积;PLR:阳性似然比;NLR:阴性似然比 -
[1] 陈红, 张重阳, 徐俊祥. 急诊院前、院内心脏骤停患者心肺复苏效果分析[J]. 河北医药, 2017, 39: 2475-2477. doi: 10.3969/j.issn.1002-7386.2017.16.019 [2] Chen J, Mei Z, Wang Y, et al. A nomogram to predict in hospital mortality in post-cardiac arrest patients: a retrospective cohort study[J]. Pol Arch Intern Med, 2023, 133: 16325 [3] Ngiam KY, Khor IW. Big data and machine learning algorithms for health-care delivery[J]. Lancet Oncol, 2019, 20: e262-e273. doi: 10.1016/S1470-2045(19)30149-4 [4] Rauschert S, Raubenheimer K, Melton PE, et al. Machine learning and clinical epigenetics: a review of challenges for diagnosis and classification[J]. Clin Epigenetics, 2020, 12: 51. doi: 10.1186/s13148-020-00842-4 [5] Lee YW, Choi JW, Shin EH. Machine learning model for predicting malaria using clinical information[J]. Comput Biol Med, 2021, 129: 104151. doi: 10.1016/j.compbiomed.2020.104151 [6] Hou N, Li M, He L, et al. Predicting 30-days mortality for MIMIC-Ⅲ patients with sepsis-3: a machine learning approach using XGboost[J]. J Transl Med, 2020, 18: 462. doi: 10.1186/s12967-020-02620-5 [7] Chu J, Leung KHB, Snobelen P, et al. Machine learning-based dispatch of drone-delivered defibrillators for out-of-hospital cardiac arrest[J]. Resuscitation, 2021, 162: 120-127. doi: 10.1016/j.resuscitation.2021.02.028 [8] 苏枫, 张少衡, 陈楠楠, 等. 基于机器学习分类判断算法构建心力衰竭疾病分期模型[J]. 中国组织工程研究, 2014, 49: 7938-7942. doi: 10.3969/j.issn.2095-4344.2014.49.012 [9] 张颖莹, 刘怡果, 赵丹, 等. 基于机器学习建立脓毒症心肾综合征患者早期死亡风险预测模型[J]. 中华肾脏病杂志, 2022, 38: 785-793. doi: 10.3760/cma.j.cn441217-20211126-00113 [10] Stevens RD. Machine Learning to Decode the Electroencephalography for Post Cardiac Arrest Neuroprognostication[J]. Crit Care Med, 2019, 47: 1474-1476. doi: 10.1097/CCM.0000000000003932 [11] Jennings JB. Can machine learning predict recurrent cardiac arrest?[J]. Resuscitation, 2023, 184: 109704. doi: 10.1016/j.resuscitation.2023.109704 [12] Blomberg SN, Folke F, Ersbøll AK, et al. Machine learning as a supportive tool to recognize cardiac arrest in emergency calls[J]. Resuscitation, 2019, 138: 322-329. doi: 10.1016/j.resuscitation.2019.01.015 [13] Kwon JM, Jeon Kh, Kim HM, et al. Deep-learning-based out-of-hospital cardiac arrest prognostic system to predict clinical outcomes[J]. Resuscitation, 2019, 139: 84-91. doi: 10.1016/j.resuscitation.2019.04.007 [14] Lundberg SM, Erion G, Chen H, et al. From Local Explanations to Global Understanding with Explainable AI for Trees[J]. Nat Mach Intell, 2020, 2: 56-67. doi: 10.1038/s42256-019-0138-9 [15] 王鑫, 廖彬, 李敏, 等. 融合LightGBM与SHAP的糖尿病预测及其特征分析方法[J]. 小型微型计算机系统, 2022, 43: 1877-1885. https://www.cnki.com.cn/Article/CJFDTOTAL-XXWX202209012.htm [16] Wu TT, Lin XQ, Mu Y, et al. Machine learning for early prediction of in-hospital cardiac arrest in patients with acute coronary syndromes[J]. Clin Cardiol, 2021, 44: 349-356. doi: 10.1002/clc.23541 [17] 郑萍, 刘宁. 机器学习应用于院外心脏骤停神经系统预后预测模型的系统评价[J]. 中国胸心血管外科临床杂志, 2022, 29: 1172-1180. https://www.cnki.com.cn/Article/CJFDTOTAL-ZXYX202209013.htm [18] 吴秋硕, 陆宗庆, 刘瑜, 等. 机器学习应用于心脏骤停早期预测模型的系统评价[J]. 中国循证医学杂志, 2021, 21: 942-952. https://www.cnki.com.cn/Article/CJFDTOTAL-ZZXZ202108011.htm [19] Mayampurath A, Hagopian R, Venable L, et al. Comparison of Machine Learning Methods for Predicting Outcomes After In-Hospital Cardiac Arrest[J]. Crit Care Med, 2022, 50: e162-e172. doi: 10.1097/CCM.0000000000005286 [20] Yan J, Xu Y, Cheng Q, et al. LightGBM: accelerated genomically designed crop breeding through ensemble learning[J]. Genome Biol, 2021, 22: 271. doi: 10.1186/s13059-021-02492-y [21] 张红莉, 李月琴, 韩磊, 等. 基于LGBM和深度神经网络的HRRP目标识别方法[J]. 探测与控制学报, 2022, 44: 97-103, 114. https://www.cnki.com.cn/Article/CJFDTOTAL-XDYX202202017.htm [22] Rufo DD, Debelee TG, Ibenthal A, et al. Diagnosis of Diabetes Mellitus Using Gradient Boosting Machine (LightGBM)[J]. Diagnostics (Basel), 2021, 11: 1714. doi: 10.3390/diagnostics11091714 [23] Ge C, Deng F, Chen W, et al. Machine learning for early prediction of sepsis-associated acute brain injury[J]. Front Med (Lausanne), 2022, 9: 962027. http://pubmed.ncbi.nlm.nih.gov/35562803/ [24] Nadolny K, Bujak K, Obremska M, et al. Glasgow Coma Scale score of more than four on admission predicts in-hospital survival in patients after out-of-hospital cardiac arrest[J]. Am J Emerg Med, 2021, 42: 90-94. [25] 蔡兰兰, 杨增强. 70例心脏骤停后复苏患者格拉斯哥昏迷评分及预后分析[J]. 内科急危重症杂志, 2018, 24: 431-433. https://www.cnki.com.cn/Article/CJFDTOTAL-NKJW201805028.htm [26] Chen YC, Hung MS, Liu CY, et al. The association of emergency department administration of sodium bicarbonate after out of hospital cardiac arrest with outcomes[J]. Am J Emerg Med, 2018, 36: 1998-2004. http://pubmed.ncbi.nlm.nih.gov/29534919/ [27] Celik T, Ozturk C, Balta S, et al. Sodium bicarbonate dilemma in patients with out-of-hospital cardiac arrest: A double-edged sword[J]. Am J Emerg Med, 2016, 34: 1314-1315. http://www.onacademic.com/detail/journal_1000038851198510_6788.html [28] Kim HJ, Park KN, Kim SH, et al. Association between the neutrophil-to-lymphocyte ratio and neurological outcomes in patients undergoing targeted temperature management after cardiac arrest[J]. J Crit Care, 2018, 47: 227-231. http://www.sciencedirect.com/science/article/pii/S0883944118303708