In-hospital mortality prediction model for patients with traumatic shock based on machine learning
-
摘要:
背景 创伤性休克(traumatic shock,TS)是导致创伤患者死亡的重要原因,关注TS患者的预后具有重要意义。 目的 构建基于入院生命体征、入院血气分析等早期指标的机器学习模型,用以预测TS患者临床结局,辅助医务人员进行医疗决策。 方法 收集解放军总医院第一医学中心急诊科2014年6月 - 2021年12月收治的TS患者信息,分为建模队列和外部验证队列两个部分,基于决策树(decision tree,DT)、逻辑回归(logistic regression,LR)、随机森林(random forest,RF)三种分类算法建立院内死亡预测模型,比较不同模型的预测效能并选择最佳模型进行外部验证。 结果 建模过程中,筛选出TS患者281例,根据患者本次出院时是否存活,分为生存组(218例)和死亡组(63例)。内部验证中,RF模型获得的受试者工作特征曲线下面积(area under receiver operator characteristic curve,AUC)最高,为0.856(95% CI:0.847 ~ 0.865),DT模型AUC为0.756(95% CI:0.740 ~ 0.772),LR模型AUC为0.801(95% CI:0.780 ~ 0.822)。RF模型的准确率(0.807)、精确率(0.886)、召回率(0.866)、F值(0.876)、AUC(0.856)等分类器性能指标均值均大于0.8。在50例外部验证集中RF模型的表现与内部验证相似,提示RF模型对TS患者死亡风险识别能力较强。通过校准曲线和决策曲线评价三种模型的校准度和临床实用性,结果提示RF模型亦优于LR和DT模型。 结论 机器学习预测模型在预测TS患者院内死亡方面具有较好的应用前景。 Abstract:Background Traumatic shock (TS) is an important cause of death in trauma patients. It is of great significance of focusing on the prognosis of TS patients. Objective To construct machine learning models based on early indicators including vital signs and arterial blood gas analysis on admission to predict clinical outcome in TS patients for medical decision-making. Methods Clinical data about TS patients admitted to the Emergency Department of the First Medical Center of Chinese PLA General Hospital from Jun. 2014 to Dec. 2021 were retrospectively collected. In-hospital mortality prediction models were constructed based on three classification algorithms: decision tree (DT), logistic regression (LR), and random forest (RF). The prediction performance of different models was compared and the best model was selected for external validation. Results In the modeling process, 281 TS patients who met the inclusion and exclusion criteria were screened, and divided into survival group (n=218) and death group (n=63) according to whether the patients were alive at discharge. In the internal validation set, the RF model obtained the highest area under receiver operator characteristic curve (AUC), which was 0.856 (95% CI: 0.847-0.865). The AUC of DT model was 0.756 (95% CI: 0.740-0.772). The AUC of LR model was 0.801 (95% CI: 0.780-0.822). The mean values of classifier performance metrics of the RF model including accuracy (0.807), precision (0.886), recall (0.866), F-value (0.876), and AUC (0.856) were all greater than 0.8. And the RF model had similar performance in the external validation set, suggesting that the RF model had a strong ability to identify the risk of death in TS patients. The calibration and clinical usefulness of the three models were evaluated by calibration curves and decision curves, suggesting that the RF model also outperformed the LR and DT models. Conclusion Prediction models based on machine learning have a promising application in predicting in-hospital mortality of TS patients. -
Key words:
- traumatic shock /
- in-hospital mortality /
- machine learning /
- prediction models /
- emergency medicine
-
表 1 TS患者不同临床结局组基线特征比较
Table 1. Comparison of baseline characteristics between the two groups of TS patients with different clinical outcomes
特征 生存组(n=218) 死亡组(n=63) χ2/Z/t值 P值 性别/(例,%) 0.230 0.631 男 174(61.92) 52(18.51) 女 44(15.66) 11(3.91) 年龄/岁 40.243 ± 15.082 44.556 ± 16.644 1.952 0.052 转运时间/[h,Md(IQR)] 5(3,9) 3(1,6) -3.665 <0.001 伤因/(例,%) 11.265 0.024 车祸伤 127(45.20) 38(13.52) 坠落伤 26(9.25) 16(5.69) 锐器伤 31(11.03) 3(1.08) 挤压伤 17(6.05) 4(1.42) 其他伤 17(6.05) 2(0.71) 伤部/(例,%) 12.116 0.033 颅脑伤 90(32.03) 45(16.01) 面颈伤 99(35.23) 47(16.73) 胸背伤 126(44.84) 46(16.37) 腰腹伤 106(37.72) 38(13.52) 会阴骨盆伤 48(17.08) 26(9.25) 四肢伤 100(35.59) 21(7.47) 体温/[℃,Md(IQR)] 36.5(36,37) 36.1(35.7,36.6) -3.560 <0.001 心率/[min-1,Md(IQR)] 121(102,140) 129(114,136) -1.638 0.101 呼吸频率/[次·min-1,Md(IQR)] 21(20,22) 20(16,22) -2.029 0.042 收缩压/mmHg 99.674 ± 21.185 85.270 ± 31.394 -3.424 0.001 舒张压/mmHg 62.872 ± 16.102 53.000 ± 23.084 -3.178 0.002 动脉血氧饱和度/[%,Md(IQR)] 98(96,98) 95(85,98) -4.719 <0.001 意识/(例,%) 50.919 <0.001 格拉斯哥昏迷评分=15 132(46.98) 6(2.14) 格拉斯哥昏迷评分<15 86(30.60) 57(20.28) 酸碱度 7.358 ± 0.088 7.230 ± 0.150 -6.458 <0.001 二氧化碳分压/[mmHg,Md(IQR)] 34(29,38) 36(30,43) -1.959 0.050 氧分压/[mmHg,Md(IQR)] 109(76.75,149.25) 84(51,160) -2.199 0.028 血钠/[mmol·L-1,Md(IQR)] 138(136,140) 139(136,141) -1.427 0.153 血钾/(mmol·L-1, Md[IQR]) 3.9(3.5,4.2) 3.7(3.4,4.3) -0.369 0.712 血钙/(mmol·L-1) 1.110 ± 0.593 1.106 ± 0.763 -0.407 0.685 血糖/[mmol·L-1,Md(IQR)] 8.9(7.5,11.4) 13.6(9.4,18.3) -5.080 <0.001 血乳酸/[mmol·L-1,Md(IQR)] 3.6(2.4,5.625) 8.3(4.4,12) -6.491 <0.001 红细胞比容/% 32.991 ± 8.788 29.492 ± 9.170 -2.756 0.006 实际碳酸氢盐/(mmol·L-1) 19.277 ± 4.267 16.279 ± 5.603 -3.929 <0.001 标准碳酸氢盐/(mmol·L-1) 20.609 ± 3.727 16.389 ± 5.142 -6.070 <0.001 二氧化碳总量/(mmol·L-1) 20.329 ± 4.425 17.479 ± 5.888 -3.562 0.001 碱剩余/(mmol·L-1) -5.483 ± 4.712 -10.475 ± 6.514 -5.669 <0.001 表 2 三种模型对TS患者院内死亡预测效能
Table 2. In-hospital mortality prediction performance of 3 models for TS patients
模型 准确率 精确率 召回率 F值 特异度 AUC (95% CI) 决策树 0.760 0.892 0.792 0.839 0.637 0.756(0.740 ~ 0.772) 逻辑回归 0.790 0.910 0.815 0.859 0.698 0.801(0.780 ~ 0.822) 随机森林 0.807 0.886 0.866 0.876 0.581 0.856(0.847 ~ 0.865) 表 3 RF对TS患者院内死亡预测效能内外部验证比较
Table 3. Comparison of in-hospital mortality prediction performance of RF models for TS patients in internal and external validation
验证集 准确率 精确率 召回率 F值 AUC 特异度 外部验证 0.800 0.892 0.846 0.868 0.741 0.636 内部验证 0.807 0.886 0.866 0.876 0.856 0.581 -
[1] Fang X,Zeng G,Linnan HW,et al. The incidence and economic burden of injuries in Jiangxi,China[J]. Public Health,2016,138: 138-145. doi: 10.1016/j.puhe.2016.03.024 [2] Zhou MG,Wang HD,Zeng XY,et al. Mortality,morbidity,and risk factors in China and its provinces,1990-2017:a systematic analysis for the Global Burden of Disease Study 2017[J]. Lancet,2019,394(10204): 1145-1158. doi: 10.1016/S0140-6736(19)30427-1 [3] 徐荣彬,温勃,宋逸,等. 1990—2016年中国青少年死亡率及主要死因变化[J]. 中华预防医学杂志,2018,52(8): 802-808. doi: 10.3760/cma.j.issn.0253-9624.2018.08.006 [4] 中国研究型医院学会卫生应急学专业委员会,中国研究型医院学会心肺复苏学专业委员会,河南省医院协会心肺复苏专业委员会. 创伤性休克急救复苏创新技术临床应用中国专家共识[J]. 实用休克杂志(中英文),2019(4): 240-245. [5] 郏瑞琪. 基于改进ANN的可解释性创伤失血性休克预测模型研究[D]. 北京: 北京交通大学, 2021. [6] Nederpelt CJ,Mokhtari AK,Alser O,et al. Development of a field artificial intelligence triage tool:confidence in the prediction of shock,transfusion,and definitive surgical therapy in patients with truncal gunshot wounds[J]. J Trauma Acute Care Surg,2021,90(6): 1054-1060. doi: 10.1097/TA.0000000000003155 [7] 邓芳帆. 便携式血气分析仪在儿科血气分析中的应用[J]. 中国医疗器械信息,2020,26(13): 158-159. doi: 10.3969/j.issn.1006-6586.2020.13.073 [8] 任珍,李姝,赵静静,等. 机器学习在急诊医学中应用的研究进展及展望[J]. 中国急救医学,2021,41(3): 261-265. doi: 10.3969/j.issn.1002-1949.2021.03.015 [9] 王乐,韩萌,李小娟,等. 不平衡数据集分类方法综述[J]. 计算机工程与应用,2021,57(22): 42-52. doi: 10.3778/j.issn.1002-8331.2107-0097 [10] Upadhyaya GK,Iyengar KP,Jain VK,et al. Evolving concepts and strategies in the management of polytrauma patients[J]. J Clin Orthop Trauma,2021,12(1): 58-65. doi: 10.1016/j.jcot.2020.10.021 [11] Gruen RL,Brohi K,Schreiber M,et al. Haemorrhage control in severely injured patients[J]. Lancet,2012,380(9847): 1099-1108. doi: 10.1016/S0140-6736(12)61224-0 [12] 张红亮,冯聪,王静,等. 损伤控制性手术在创伤治疗中的应用综述[J]. 解放军医学院学报,2020,41(12): 1262-1264. doi: 10.3969/j.issn.2095-5227.2020.12.020 [13] 赵鹏跃,刘逸尘,许晓蕾,等. 创伤失血性休克早期液体复苏研究进展[J]. 解放军医学院学报,2019,40(10): 985-987. doi: 10.3969/j.issn.2095-5227.2019.10.017 [14] 吴京兰,徐红波. 创伤性休克的早期识别与处理[J]. 实用休克杂志(中英文),2019,3(2): 81-85. [15] Anderson MW,Watson GA. Traumatic shock:the fifth shock[J]. J Trauma Nurs,2013,20(1): 37-43. doi: 10.1097/JTN.0b013e318286620a [16] Falzone E,Pasquier P,Hoffmann C,et al. Triage in military settings[J]. Anaesth Crit Care Pain Med,2017,36(1): 43-51. doi: 10.1016/j.accpm.2016.05.004 [17] Jiang XB,Jiang P,Mao YS. Performance of Modified Early Warning Score (MEWS) and Circulation,Respiration,Abdomen,Motor,and Speech (CRAMS) score in trauma severity and in-hospital mortality prediction in multiple trauma patients:a comparison study[J]. PeerJ,2019,7: e7227. doi: 10.7717/peerj.7227 [18] Raux M,Sartorius D,le Manach Y,et al. What do prehospital trauma scores predict besides mortality?[J]. J Trauma,2011,71(3): 754-759. [19] 雷明星,刘蜀彬. 创伤评分系统的研究进展[J]. 灾害医学与救援(电子版),2017,6(1): 60-63. [20] 唐瑞,唐雯,王导新. 机器学习对创伤合并急性呼吸窘迫综合征患者院内死亡的预测价值[J]. 中华危重病急救医学,2022,34(3): 260-264. doi: 10.3760/cma.j.cn121430-20211117-01741 [21] 赵宇卓. 基于创伤大数据的创伤失血性休克伤情预测预警模型建立研究[D]. 北京: 中国人民解放军医学院, 2019. [22] Li KY, Wu HT, Pan F, et al. A machine learning-based model to predict acute traumatic coagulopathy in trauma patients upon emergency hospitalization[J]. Clin Appl Thromb Hemost, 2020, 26: 1076029619897827. -