切换至 "中华医学电子期刊资源库"

中华实验和临床感染病杂志(电子版) ›› 2025, Vol. 19 ›› Issue (06) : 368 -377. doi: 10.3877/cma.j.issn.1674-1358.2025.06.007

论著

重症烧伤患者多重耐药菌感染风险的机器学习预测模型构建及验证
李东晖()   
  1. 100020 北京,首都医科大学附属朝阳医院石景山院区SICU
  • 收稿日期:2024-12-24 出版日期:2025-12-15
  • 通信作者: 李东晖

Construction and verification of machine learning prediction model for risk of multiple drug-resistant bacterial infection in patients with severe burn

Donghui Li()   

  1. SICU Shijingshan Campus, Beijing Chao-yang Hospital, Capital Medical University, Beijing 100020, China
  • Received:2024-12-24 Published:2025-12-15
  • Corresponding author: Donghui Li
引用本文:

李东晖. 重症烧伤患者多重耐药菌感染风险的机器学习预测模型构建及验证[J/OL]. 中华实验和临床感染病杂志(电子版), 2025, 19(06): 368-377.

Donghui Li. Construction and verification of machine learning prediction model for risk of multiple drug-resistant bacterial infection in patients with severe burn[J/OL]. Chinese Journal of Experimental and Clinical Infectious Diseases(Electronic Edition), 2025, 19(06): 368-377.

目的

基于机器学习算法,构建重症烧伤患者多重耐药菌(MDRO)感染风险的预测模型并进行验证。

方法

收集2017年5月至2024年6月首都医科大学附属朝阳医院收治的675例重症烧伤患者的临床资料。根据是否发生MDRO感染,将患者分为非MDRO感染组(535例)与MDRO感染组(140例)。通过最小绝对收缩和选择算子(LASSO)筛选发生MDRO感染的变量因素。将入选样本量按照7∶3比例随机分为训练集(473例)与验证集(202例),分别构建极限梯度提升树(XGBoost)、Logistic回归、支持向量机(SVM)、K近邻(KNN)和高斯朴素贝叶斯分类(GNB)预测模型。采用受试者工作特征(ROC)曲线下面积(AUC)、决策曲线分析(DCA)和校准曲线评估各模型的性能。使用Shapley加法解释(SHAP)分析每个变量对预测模型的影响。

结果

675例患者中发生MDRO感染者140例,发生率为20.74%(140/675)。非MDRO感染组与MDRO感染组患者烧伤深度(χ2=32.244、P<0.001)、烧伤面积(χ2=9.438、P=0.002)、低蛋白血症(χ2=4.713、P=0.030)、入重症监护室(ICU)前使用抗菌药物(χ2=11.073、P=0.001)、机械通气(χ2=4.125、P=0.042)、抗菌药物使用≥3种(χ2=22.031、P<0.001)、抗菌药物使用时间≥7 d(χ2=10.377、P=0.001)、留置导尿管(χ2=13.519、P<0.001)、深静脉置管(χ2=11.743、P<0.001)、住ICU时间(Z=21.359、P<0.001)与住院时间(Z=11.271、P<0.001)差异均有统计学意义。基于LASSO回归筛选MDRO感染的变量因素构建5种预测模型,其中GNB模型综合表现最好,训练集AUC为0.877(95%CI:0.837~0.917),验证集AUC为0.872(95%CI:0.809~0.934),校准曲线显示模型预测概率与实际结果最相近,DCA曲线显示,风险概率值为0.10~1.00时,临床净收益高。SHAP分析显示,按影响重症烧伤患者发生MDRO感染因素的重要性排序分别为住ICU时间、住院时间、烧伤深度、抗菌药物使用时间≥7 d、留置导尿管、深静脉置管以及入ICU前使用抗菌药物。

结论

以烧伤深度、抗菌药物使用时间≥7 d、留置导尿管、深静脉置管、入ICU前使用抗菌药物、住ICU时间、住院时间构建的GNB模型为重症烧伤患者发生MDRO感染的最佳预测模型,对医护人员制定有效防控策略有一定的指导意义。

Objective

To construct and validate a prediction model of multidrug-resistant organism (MDRO) infection risk in severe burned patients based on machine learning algorithm.

Methods

Clinical data of 675 patients with severe burn treated in Beijing Chao-yang Hospital, Capital Medical University from May 2017 to June 2024 were collected. Patients were divided into non-MDRO infection group (535 cases) and MDRO infection group (140 cases) based on whether MDRO infection occurred. The variable factors for MDRO infection were screened by the least absolute shrinkage and selection operator (LASSO). All the sample sizes were randomly divided into the training set (473 cases) and verification set (202 cases) according to the ratio of 7∶3, and the extreme Gradient Boost tree (XGBoost), Logistic regression, support vector machine (SVM), K-nearest neighbor (KNN) and Gauss Naive Bayes classification (GNB) prediction model were constructed, respectively. The performance of each model was evaluated by area under the receiver operating characteristic (ROC) curve (AUC), decision curve analysis (DCA) and calibration curve. The influence of each variable factor on the prediction model was analyzed using Shapley additive interpretation (SHAP).

Results

Among the 675 patients, 140 cases developed MDRO infection, with an incidence of 20.74% (140/675). The burn depth (χ2=32.244, P<0.001), burn area (χ2=9.438, P=0.002), hypoalbuminemia (χ2=4.713, P=0.030), usage of antibacterial drugs before intensive care unit (ICU) admission (χ2=11.073, P=0.001), mechanical ventilation (χ2=4.125, P=0.042), usage of≥3 antibacterial drugs (χ2=22.031, P<0.001), duration of antibiotic use≥7 days (χ2=10.377, P=0.001), indwelling catheter (χ2=13.519, P<0.001), deep vein catheterization (χ2=11.743, P<0.001), duration of ICU stay (Z=21.359, P<0.001) and duration of hospitalization (Z=11.271, P<0.001) between patients in non-MDRO infection group and MDRO infection group were all with significant differences. Five prediction models were constructed based on LASSO regression screening variable factors of MDRO infection, among which the GNB model had the best overall performance, with the AUC of 0.877 (95%CI: 0.837-0.917) for the training set and 0.872 (95%CI: 0.809-0.934) for the validation set, and the calibration curve showed that the model’s predicted probability was closest to the actual results, and the DCA curve showed that when the risk probability value was between 0.10-1.00, the clinical net benefit was high. SHAP analysis showed that the factors affecting the occurrence of MDRO infection in severe burned patients, ranked by importance, were as follows: duration of ICU stay, duration of hospitalization, depth of burn, duration of antibiotic use≥7 days, indwelling urinary catheter, deep vein catheterization and use of antibiotics before ICU admission.

Conclusions

The GNB model based on burn depth, duration of antibiotic use≥7 days, indentation catheter, deep vein catheterization, use of antibacterial drugs before admission to ICU, duration of stay in ICU and duration of hospitalization are the best prediction models for MDRO infection in patients with severe burns, which has some guiding significance for clinicians to develop effective prevention and control strategies.

表1 非MDRO感染组与MDRO感染组重症烧伤患者的一般资料
一般资料 非MDRO感染组(535例) MDRO感染组(140例) 统计量 P
性别 [例(%)] χ2=0.818 0.366a
345(64.49) 96(68.57)
190(35.51) 44(31.43)
年龄 [例(%)] χ2=0.354 0.552a
≥40岁 283(52.90) 78(55.71)
<40岁 252(47.10) 62(44.29)
BMI(
±s,kg/m2
24.15±3.54 24.22±3.75 t=0.206 0.837
吸烟 [例(%)] χ2=1.117 0.291a
158(29.53) 35(25.00)
377(70.47) 105(75.00)
高血压 [例(%)] χ2=0.016 0.900a
169(31.59) 45(32.14)
366(68.41) 95(67.86)
糖尿病 [例(%)] χ2=4.306 0.038a
133(24.86) 47(33.57)
402(75.14) 93(66.43)
冠心病 [例(%)] χ2=0.442 0.506a
109(20.37) 25(17.86)
426(79.63) 115(82.14)
慢性肺部疾病 [例(%)] χ2=0.614 0.433a
99(18.50) 30(21.43)
436(81.50) 110(78.57)
烧伤类型 [例(%)] χ2=0.444 0.505a
火焰 434(81.12) 117(83.57)
其他 101(18.18) 23(16.43)
烧伤深度 [例(%)] χ2=32.244 <0.001a
Ⅰ~浅Ⅱ 378(70.65) 63(45.00)
深Ⅱ~Ⅲ 157(29.35) 77(55.00)
烧伤面积 [例(%)] χ2=9.438 0.002a
<50% 307(57.38) 60(42.86)
≥50% 228(42.62) 80(57.14)
低蛋白血症 [例(%)] χ2=4.713 0.030a
321(60.00) 98(70.00)
214(40.00) 42(30.00)
气管切开 [例(%)] χ2=1.296 0.255a
110(20.56) 35(25.00)
425(79.44) 105(75.00)
入ICU前使用抗菌药物 [例(%)] χ2=11.073 0.001a
182(34.02) 69(49.29)
353(65.98) 71(50.71)
机械通气 [例(%)] χ2=4.125 0.042a
162(30.28) 55(39.29)
373(69.72) 85(60.71)
抗菌药物使用≥ 3种 [例(%)] χ2=22.031 <0.001a
225(42.06) 90(64.29)
310(57.94) 50(35.71)
抗菌药物使用时间 [例(%)] χ2=10.377 0.001a
≥7 d 188(35.14) 70(50.00)
<7 d 347(64.85) 70(50.00)
使用糖皮质激素 [例(%)] χ2=0.178 0.673a
196(36.64) 54(38.57)
339(63.36) 86(61.43)
留置导尿管 [例(%)] χ2=13.519 <0.001a
405(75.70) 126(90.00)
130(24.30) 14(10.00)
深静脉置管 [例(%)] χ2=11.743 <0.001a
317(59.25) 105(75.00)
218(40.75) 35(25.00)
胃管插管 [例(%)] χ2=1.205 0.272a
368(68.79) 103(73.57)
167(31.21) 37(26.43)
住ICU时间 [MP25P75),d] 10(5,20) 15(8,26) Z=21.359 <0.001
住院时间[MP25P75),d] 22(17,32) 28(18,43) Z=11.271 <0.001
图1 LASSO回归分析筛选变量 注:A:变量LASSO系数分布图;B:LASSO回归变量收缩图
表2 预测重症烧伤患者发生MDRO感染风险的各模型参数设置
图2 预测重症烧伤患者发生MDRO感染风险的各模型曲线图
表3 预测重症烧伤患者发生MDRO感染风险的各模型预测性能
图3 GNB模型训练、验证与测试
图4 构建GNB模型的SHAP分析图 注:A:SHAP图;B:变量重要性排序
[1]
袁志强, 彭毅志. 烧伤重症监护病房多重耐药菌感染的应对策略及思考[J]. 中华烧伤杂志,2021,37(6):524-529.
[2]
He Q, Chen C, Gao S, et al. Predictive value of perioperative peripheral blood cells counts for bacteremia and 90-day mortality in severe burn patients[J]. Burns,2023,49(6):1412-1421.
[3]
Lachiewicz AM, Hauck CG, Weber DJ, et al. Bacterial infections after burn injuries: impact of multidrug resistance[J]. Clin Infect Dis,2017,65(12):2130-2136.
[4]
Vickers ML, Dulhunty JM, Ballard E, et al. Risk factors for multidrug-resistant Gram-negative infection in burn patients[J]. ANZ J Surg,2018,8(5):480-485.
[5]
尹燕燕, 刘爱贤. 多重耐药菌感染呼吸机相关性肺炎的危险因素及预后分析[J/CD]. 中华实验和临床感染病杂志(电子版),2024,18(2):83-90.
[6]
Han SB, Jeon K, Kim M, et al. Risk factors for in-hospital acquisition of carbapenem-resistant Enterobacterales in patients with severe burns[J]. J Hosp Infect,2024,149(1):165-171.
[7]
叶楠, 祝闯, 徐丰博, 等. 机器学习算法构建急性心肌梗死患者发生急性肾损伤风险预测模型并与传统模型比较[J]. 中华肾脏病杂志,2024,40(3):175-182.
[8]
王亚喜, 张玲慧, 曹文, 等. 重症监护室病人多重耐药菌感染预测模型的构建[J]. 青岛大学学报(医学版),2024,60(1):115-119.
[9]
中华人民共和国卫生部. 医院感染诊断标准(试行)[S]. 2001. 
[10]
中华人民共和国卫生部. 多重耐药菌医院感染预防与控制技术指南(试行)[J]. 中国危重病急救医学,2011,23(2):69-71.
[11]
Cleland H, Stewardson A, Padiglione A, et al. Bloodstream infections and multidrug resistant bacteria acquisition among burns patients in Australia and New Zealand: A registry-based study[J]. Burns,2024,50(6):1544-1554.
[12]
彭源, 傅跃先. 多药耐药细菌感染的治疗研究进展[J]. 中华烧伤杂志,2016,32(9):539-541.
[13]
滕支梅, 纪赓, 黄劲华, 等. 2018-2020年泰州市某医院烧伤科患者多重耐药菌感染危险因素分析[J]. 东南大学学报(医学版),2022,41(6):812-817.
[14]
Elsheikh R, Makram AM. Multidrug-resistant organisms: The silent plight of burn patients[J]. J Burn Care Res,2024,45(4):877-886.
[15]
吴红, 蔡伟茶, 金琪琪, 等. 髋部骨折老年患者术后1年死亡的危险因素及其预测模型的准确性: 基于LASSO-logistic回归[J]. 中华麻醉学杂志,2024,44(1):15-19.
[16]
Buriro F, Ishaque S, Saeed A, et al. Prevalence of multidrug-resistant organism in ICU burns patients at tertiary care hospital[J]. J Burn Care Res,2023,44(4):949-954.
[17]
van Duin D, Strassle PD, DiBiase LM, et al. Timeline of health care-associated infections and pathogens after burn injuries[J]. Am J Infect Control,2016,44(12):1511-1516.
[18]
谭芬叶, 赵蓓, 陈亚男, 等. 神经内科重症监护室多重耐药菌感染预测模型构建[J]. 中华医院感染学杂志,2024,34(10):1575-1578.
[19]
吴杨炀, 曹玲, 仲悦萍, 等. 烧伤合并多重耐药革兰阴性菌感染的调查及院内感染发生危险因素的logistic分析[J]. 中国卫生统计,2021,38(3):396-398.
[20]
Sadaka MS, Abdeldaim DE. Study of the relationship between patient demographics, burn etiology, and the incidence of burn wound infection in Tanta University Burn Unit[J]. Ann Burns Fire Disasters,2024,37(1):3-9.
[21]
Ilyas F, James A, Khan S, et al. Multidrug-resistant pathogens in wound infections: A systematic review[J]. Cureus,2024,16(4):1-9.
[22]
卢蔚薇, 赖宇涛, 童雅婵, 等. 新生儿重症监护病房多重耐药菌反复感染的危险因素分析[J]. 传染病信息,2023,36(3):243-247.
[23]
桑玉还, 严忠婷, 袁媛, 等. 重症监护病房患者多重耐药菌感染风险预测模型的系统评价[J]. 中国感染控制杂志,2023,22(4):442-450.
[24]
沈丽, 南海英, 黄陆光, 等. 烧伤重症监护病房住院患者多重耐药菌感染及其危险因素分析[J]. 中国消毒学杂志,2024,41(4):287-289.
[25]
张艳霞, 姚新宝, 热娜·吐尔逊. 480例烧伤科手术患者医院感染危险因素调查[J]. 中国感染控制杂志,2016,15(12):926-929.
[26]
刘娟, 张小霞, 臧凤, 等. 老年患者导管相关性尿路多重耐药菌感染危险因素及其预测模型[J]. 重庆医学,2024,53(10):1519-1524.
[1] 张亚琼, 张雨霖, 丁科, 张晓霞, 胡文佳, 陈铁龙, 宋世会, 熊勇. 人类免疫缺陷病毒感染/获得性免疫缺陷综合征患者接受抗反转录病毒治疗四年后发生免疫重建不全的风险预测模型[J/OL]. 中华实验和临床感染病杂志(电子版), 2025, 19(06): 335-344.
[2] 钱何布, 朱林, 姚月平, 姚峰, 李玉卓, 马家驹, 晏倩, 倪晓艳. 应用机器学习建立脓毒性休克患者住院28天死亡预测模型及验证[J/OL]. 中华实验和临床感染病杂志(电子版), 2025, 19(05): 288-297.
[3] 徐世伟, 廖杜荣, 张镐, 叶辉, 陈志平, 雒洪志. 基于一种新炎症-营养指标构建结直肠癌术前淋巴结转移预测模型[J/OL]. 中华普通外科学文献(电子版), 2025, 19(06): 383-389.
[4] 孙圣梅, 习一清, 安宁. 人表皮生长因子受体2阳性型乳腺癌新辅助治疗响应的基因预测模型[J/OL]. 中华普通外科学文献(电子版), 2025, 19(05): 332-339.
[5] 罗仲燃, 曾智豪, 黄梦娟, 何晓艺. 乳腺癌术后腋窝淋巴结负荷的多因素分析及预测模型的建立及验证[J/OL]. 中华普外科手术学杂志(电子版), 2026, 20(01): 46-50.
[6] 谷瑞辰, 林福杨, 刘永达. 术前尿培养阴性患者PCNL术后尿路感染的预测模型构建[J/OL]. 中华腔镜泌尿外科杂志(电子版), 2025, 19(06): 772-777.
[7] 黄少坚, 梁汉标, 李清平, 唐善华, 李青妍, 李芷西, 黄灿, 王小振, 陈灿辉, 王恺, 李川江. 基于影像组学和临床特征构建肝癌新辅助/转化治疗后病理学完全缓解预测模型[J/OL]. 中华肝脏外科手术学电子杂志, 2025, 14(06): 860-867.
[8] 林水荣, 宋子敏, 于玺, 李绍强, 华赟鹏, 沈顺利. 术前抗病毒治疗对HBV相关肝癌肝切除术后肝衰竭影响[J/OL]. 中华肝脏外科手术学电子杂志, 2025, 14(05): 700-706.
[9] 赵俊宇, 林航宇, 李会灵, 王显飞, 游川. 肝癌肝切除术后大量腹水预测模型的建立与验证[J/OL]. 中华肝脏外科手术学电子杂志, 2025, 14(05): 740-747.
[10] 刘燕, 马亦旻. 消化道早癌患者内镜黏膜下剥离术后局部复发预测模型构建与验证[J/OL]. 中华消化病与影像杂志(电子版), 2025, 15(06): 576-582.
[11] 李曰平, 鞠倩, 张汝梦, 韩博. 基于CT影像组学预测胃癌根治术复发风险的临床研究[J/OL]. 中华消化病与影像杂志(电子版), 2025, 15(05): 460-466.
[12] 李姁, 冯中秋. HBV相关原发性肝癌患者经导管动脉化疗栓塞术后严重腹痛预测模型的构建及验证[J/OL]. 中华消化病与影像杂志(电子版), 2025, 15(05): 480-485.
[13] 王超, 张晓会, 李晓帆, 赵海丹. 维持性血液透析患者感染与心脑血管死亡风险的比较及联合预测模型构建[J/OL]. 中华临床医师杂志(电子版), 2025, 19(09): 675-681.
[14] 郭岩, 赵灵芝, 石光英. 代谢相关脂肪性肝病患者发生冠心病的风险预测模型[J/OL]. 中华临床医师杂志(电子版), 2025, 19(09): 682-688.
[15] 皇立媛, 浦洁, 王苏贵, 陈婷婷, 朱德慧, 胡雪. 中青年脑卒中患者应激障碍风险预测模型的构建与验证[J/OL]. 中华临床医师杂志(电子版), 2025, 19(07): 504-512.
阅读次数
全文


摘要


AI


AI小编
你好!我是《中华医学电子期刊资源库》AI小编,有什么可以帮您的吗?