Email Alert | RSS

北京大学学报（医学版） ›› 2021, Vol. 53 ›› Issue (6): 1163-1170. doi: 10.19723/j.issn.1671-167X.2021.06.026

• 论著 • 上一篇    下一篇

基于机器学习的重症监护室超长入住时长预测

吴静依1,2,林瑜1,蔺轲1,胡永华1,3,孔桂兰2,4,△()   

  1. 1.北京大学公共卫生学院流行病与卫生统计学系,北京 100191
    2.浙江省北大信息技术高等研究院,杭州 311200
    3.北京大学医学信息学中心,北京 100191
    4.北京大学健康医疗大数据国家研究院,北京 100191
  • 收稿日期:2019-11-26 出版日期:2021-12-18 发布日期:2021-12-13
  • 通讯作者: 孔桂兰 E-mail:guilan.kong@hsc.pku.edu.cn
  • 基金资助:
    国家自然科学基金(81771938);国家自然科学基金(91846101);北京市自然科学基金(7212201);科技创新2030项目(2018AAA0102100);北京大学医学部-密歇根大学医学院转化医学与临床研究联合研究所项目(BMU2020JI011)

Predicting prolonged length of intensive care unit stay via machine learning

WU Jing-yi1,2,LIN Yu1,LIN Ke1,HU Yong-hua1,3,KONG Gui-lan2,4,△()   

  1. 1. Department of Epidemiology and Biostatistics, Peking University School of Public Health, Beijing 100191, China
    2. Advanced Institute of Information Technology, Peking University, Hangzhou 311200, China
    3. Peking University Medical Informatics Center, Beijing 100191, China
    4. National Institute of Health Data Science, Peking University, Beijing 100191, China
  • Received:2019-11-26 Online:2021-12-18 Published:2021-12-13
  • Contact: Gui-lan KONG E-mail:guilan.kong@hsc.pku.edu.cn
  • Supported by:
    National Natural Science Foundation of China(81771938);National Natural Science Foundation of China(91846101);Beijing Municipal Science & Technology Commission(7212201);Chinese Scientific and Technical Innovation Project 2030(2018AAA0102100);University of Michigan Health System-Peking University Health Science Center Joint Institute for Translational and Clinical Research(BMU2020JI011)

RICH HTML

   

PDF (PC)

摘要：

目的:基于三种机器学习算法——支持向量机(support vector machine,SVM)、分类回归树(classification and regression tree,CART)和随机森林(random forest,RF),构建重症监护室(intensive care unit,ICU)患者的ICU入住时长(length of ICU stay,LOS-ICU)分类预测模型,并与传统的定制版简化急性生理功能评分Ⅱ(simplified acute physiology score Ⅱ,SAPS-Ⅱ)模型进行比较。方法:使用美国大型重症医疗数据库(medical information mart for intensive care Ⅲ,MIMIC-Ⅲ),以ICU患者是否发生超长LOS-ICU(prolonged LOS-ICU,pLOS-ICU)作为结局指标,构建定制版SAPS-Ⅱ、SVM、CART和RF模型,使用递归特征消除法进行特征选择,基于五折交叉验证找出最佳预测模型。模型的预测性能评价指标包括Brier评分、受试者工作特征(receiver operation characteristic,ROC)曲线下面积(area under the ROC curve,AUROC)和估计校准度指数(estimated calibration index,ECI),模型性能指标之间的比较使用双侧t检验。使用本研究中预测性能最好的模型识别出来的各预测变量重要性排序结果,给出重要性排序前五位的预测变量。结果:最终共纳入40 200例ICU患者,发生pLOS-ICU的患者23.7%。其中,男性患者57.6%,患者平均年龄为(61.9±16.5)岁。五折交叉验证结果显示,相比于定制版SAPS-Ⅱ模型,三种机器学习模型的预测性能在各个指标上均有明显提升,且差异均具有统计学意义(P<0.01)。其中,RF模型在综合预测性能、区分度与校准度三个方面均表现最优,其Brier评分、AUROC和ECI分别为0.145、0.770和7.259。校准曲线结果显示,在高pLOS-ICU发生风险的ICU人群中,RF模型倾向于略微高估其风险;在低pLOS-ICU发生风险的ICU人群中,RF模型倾向于略微低估其风险。基于性能最优的RF模型识别的对pLOS-ICU预测最重要的五个变量依次为年龄、心率、收缩压、体温和动脉血氧分压与吸入氧分数之比。结论:基于机器学习方法构建ICU患者的pLOS-ICU预测模型相比于传统的定制版SAPS-Ⅱ模型,预测性能均有明显提升,其中,基于RF方法的pLOS-ICU预测模型性能最优,具有很大的临床应用潜力。

关键词: 重症监护室, 住院时长, 机器学习, 随机森林, 简化急性生理功能评分

Abstract:

Objective: To construct length of intensive care unit (ICU) stay (LOS-ICU) prediction models for ICU patients, based on three machine learning models: support vector machine (SVM), classification and regression tree (CART), and random forest (RF), and to compare the prediction perfor-mance of the three machine learning models with the customized simplified acute physiology score Ⅱ(SAPS-Ⅱ) model. Methods: We used medical information mart for intensive care (MIMIC)-Ⅲ database for model development and validation. The primary outcome was prolonged LOS-ICU(pLOS-ICU), defined as longer than the third quartile of patients’ LOS-ICU in the studied dataset. The recursive feature elimination method was used to do feature selection for three machine learning models. We utilized 5-fold cross validation to evaluate model prediction performance. The Brier value, area under the receiver operation characteristic curve (AUROC), and estimated calibration index (ECI) were used as perfor-mance measures. Performances of the four models were compared, and performance differences between the models were assessed using two-sided t test. The model with the best prediction performance was employed to generate variable importance ranking, and the identified top five important predictors were pre-sented. Results: The final cohort in our study consisted of 40 200 eligible ICU patients, of whom 23.7% were with pLOS-ICU. The proportion of the male patients was 57.6%, and the age of all the ICU patients was (61.9±16.5) years. Results showed that the three machine learning models outperformed the customized SAPS-Ⅱ model in terms of all the performance measures with statistical significance (P<0.01). Among the three machine learning models, the RF model achieved the best overall performance (Brier value, 0.145), discrimination (AUROC, 0.770) and calibration (ECI, 7.259). The calibration curve showed that the RF model slightly overestimated the risk of pLOS-ICU in high-risk ICU patients, but underestimated the risk of pLOS-ICU in low-risk ICU patients. Top five important predictors for pLOS-ICU identified by the RF model included age, heart rate, systolic blood pressure, body tempe-rature, and ratio of arterial oxygen tension to the fraction of inspired oxygen(PaO2/FiO2). Conclusion: The RF algorithm-based pLOS-ICU prediction model had a best prediction performance in this study. It lays a foundation for future application of the RF-based pLOS-ICU prediction model in ICU clinical practice.

Key words: Intensive care units, Length of stay, Machine learning, Random forest, Simplified acute physiology score

中图分类号: 

  • R181.2

表1

MIMIC-Ⅲ数据库中ICU患者的基本特征(n=40 200)"

Items Data
Age/years 61.9±16.5
Gender, n (%)
Male 23 145 (57.6)
Female 17 055 (42.4)
SAPS-Ⅱ score 32.5 ± 12.6
LOS-ICU / d, M (Q1-Q3) 2.1 (1.2-3.9)
pLOS-ICU, n (%) 9 514 (23.7)

图1

ICU患者pLOS-ICU发生率依年龄、性别分层的金字塔图"

图2

建模特征的个数与RF模型性能的关系曲线"

表2

四种模型的预测性能比较"

Model Brier value AUROC ECI
Customized SAPS-Ⅱ 0.169 0.666 8.435
SVM 0.152* 0.738* 7.596*
CART 0.154* 0.730* 7.712*
RF 0.145* 0.770* 7.259*

图3

RF模型的校准曲线"

[1] Bin D, Youzhong A, Yan K, et al. Characteristics of critically ill patients in ICUs in mainland China[J]. Crit Care Med, 2013, 41(1):84-92.
doi: 10.1097/CCM.0b013e31826a4082 pmid: 23222268
[2] Milbrandt EB, Kersten A, Rahim MT, et al. Growth of intensive care unit resource use and its estimated cost in Medicare[J]. Crit Care Med, 2008, 36(9):2504-2510.
doi: 10.1097/CCM.0b013e318183ef84 pmid: 18679127
[3] Tsai PF, Chen PC, Chen YY, et al. Length of hospital stay prediction at the admission stage for cardiology patients using artificial neural network[J]. J Healthc Eng, 2016, 2016:7035463.
[4] Nassar AP, Caruso P. ICU physicians are unable to accurately predict length of stay at admission: A prospective study[J]. Int J Qual Health Care, 2016, 28(1):99-103.
doi: 10.1093/intqhc/mzv112
[5] Le Gall JR, Lemeshow S, Saulnier F. A new simplified acute physiology ccore (SAPSⅡ) based on a European/North American multicenter study[J]. JAMA, 1993, 270(24):2957-2963.
doi: 10.1001/jama.1993.03510240069035
[6] Zimmerman JE, Kramer AA, Mcnair DS, et al. Acute physiology and chronic health evaluation (APACHE) Ⅳ: Hospital mortality assessment for today’s critically ill patients[J]. Crit Care Med, 2006, 34(5):1297-1310.
pmid: 16540951
[7] Vasilevskis EE, Kuzniewicz MW, Cason BA, et al. Mortality probability model Ⅲ and simplified acute physiology scoreⅡassessing their value in predicting length of stay and comparison to APACHE Ⅳ[J]. Chest, 2009, 136(1):89-101.
doi: S0012-3692(09)60409-8 pmid: 19363210
[8] Zimmerman JE, Kramer AA, Mcnair DS, et al. Intensive care unit length of stay: Benchmarking based on acute physiology and chronic health evaluation (APACHE) Ⅳ[J]. Crit Care Med, 2006, 34(10):2517-2529.
pmid: 16932234
[9] Lin K, Hu Y, Kong G. Predicting in-hospital mortality of patients with acute kidney injury in the ICU using random forest model[J]. Int J Med Inform, 2019, 125:55-61.
doi: S1386-5056(18)31087-6 pmid: 30914181
[10] Saria S, Rajani AK, Gould J, et al. Integration of early physiolo-gical responses predicts later illness severity in preterm infants[J]. Sci Transl Med, 2010, 2(48):48-65.
[11] Johnson AEW, Pollard TJ, Shen L, et al. MIMIC-Ⅲ, a freely accessible critical care database[J]. Sci Data, 2016, 3:160035.
doi: 10.1038/sdata.2016.35
[12] Agrawal S, Chen L, Tergas A I, et al. Characteristics associated with prolonged length of stay after hysterectomy for benign gynecologic conditions [J]. Am J Obstet Gynecol, 2018, 219(1): 89.e1-89.e15.
doi: 10.1016/j.ajog.2018.05.001
[13] Wolff J, Mccrone P, Patel A, et al. Predictors of length of stay in psychiatry: Analyses of electronic medical records[J]. BMC Psychiatry, 2015, 15:238.
doi: 10.1186/s12888-015-0623-6 pmid: 26446584
[14] Louw N, Steel SJ. Variable selection in kernel Fisher discriminant analysis by means of recursive feature elimination[J]. Comput Stat Data Anal, 2006, 51(3):2043-2055.
doi: 10.1016/j.csda.2005.12.018
[15] Boser BE, Guyon IM, Vapnik VN. A training algorithm for optimal margin classifiers[C]// Proceedings of the fifth annual workshop on computational learning theory. New York: Association for Computing Machinery, 1992: 144-152.
[16] Furey TS, Cristianini N, Duffy N, et al. Support vector machine classification and validation of cancer tissue samples using microarray expression data[J]. Bioinformatics, 2000, 16(10):906-914.
pmid: 11120680
[17] Breiman L, Friedman JH, Olshen RA, et al. Classification and regression trees[M]. Spain: Routledge, 2017.
[18] Austin PC. A comparison of regression trees, logistic regression, generalized additive models, and multivariate adaptive regression splines for predicting AMI mortality[J]. Stat Med, 2007, 26(15):2937-2957.
pmid: 17186501
[19] Breiman L. Random forests[J]. Machine Learning, 2001, 45(1):5-32.
doi: 10.1023/A:1010933404324
[20] Supatcha L, Chinae T, Chakarida N, et al. Heterogeneous ensemble approach with discriminative features and modified-SMOTE bagging for pre-miRNA classification[J]. Nucleic Acids Res, 2013, 41(1):e21.
doi: 10.1093/nar/gks878
[21] Wang SQ, Yang J, Chou KC. Using stacked generalization to predict membrane protein types based on pseudo-amino acid composition[J]. J Theor Biol, 2006, 242(4):941-946.
doi: 10.1016/j.jtbi.2006.05.006
[22] Touw WG, Bayjanov JR, Overmars L, et al. Data mining in the life sciences with random forest: A walk in the park or lost in the jungle?[J]. Brief Bioinformatics, 2013, 14(3):315-326.
doi: 10.1093/bib/bbs034
[23] Steyerberg EW, Vickers AJ, Cook NR, et al. Assessing the performance of prediction models a framework for traditional and novel measures[J]. Epidemiology, 2010, 21(1):128-138.
doi: 10.1097/EDE.0b013e3181c30fb2 pmid: 20010215
[24] van Hoorde K, van Huffel S, Timmerman D, et al. A spline-based tool to assess and visualize the calibration of multiclass risk predictions[J]. J Biomed Inform, 2015, 54:283-293.
doi: 10.1016/j.jbi.2014.12.016 pmid: 25579635
[25] Peigne V, Somme D, Guerot E, et al. Treatment intensity, age and outcome in medical ICU patients: Results of a French admi-nistrative database[J]. Ann Intensive Care, 2016, 6(1):7.
doi: 10.1186/s13613-016-0107-y
[26] Fallenius M, Skrifvars MB, Reinikainen M, et al. Common intensive care scoring systems do not outperform age and glasgow coma scale score in predicting mid-term mortality in patients with spontaneous intracerebral hemorrhage treated in the intensive care unit[J]. Scand J Trauma Resusc Emerg Med, 2017, 25(1):102.
doi: 10.1186/s13049-017-0448-z
[27] Szalados JE. Long-term mortality and quality of life after prolonged mechanical ventilation. Age and functional status as determinants of intensive care unit outcome: Sound basis for health policy or tip of the outcomes iceberg[J]. Crit Care Med, 2004, 32(1):291-293.
pmid: 14707597
[28] Brandberg C, Blomqvist H, Jirwe M. What is the importance of age on treatment of the elderly in the intensive care unit?[J]. Acta Anaesthesiol Scand, 2013, 57(6):698-703.
doi: 10.1111/aas.12073 pmid: 23373851
[29] Diringer MN, Reaven NL, Funk SE, et al. Elevated body temperature independently contributes to increased length of stay in neurologic intensive care unit patients[J]. Crit Care Med, 2004, 32(7):1489-1495.
pmid: 15241093
[30] Yien HW, Hseu SS, Lee LC, et al. Spectral analysis of systemic arterial pressure and heart rate signals as a prognostic tool for the prediction of patient outcome in the intensive care unit[J]. Crit Care Med, 1997, 25(2):258-266.
pmid: 9034261
[31] Grander W, Mullauer K, Koller B, et al. Heart rate before ICU discharge: a simple and readily available predictor of short- and long-term mortality from critical illness[J]. Clin Res Cardiol, 2013, 102(8):599-606.
doi: 10.1007/s00392-013-0571-4
[32] Esteve F, Lopez-Delgado JC, Javierre C, et al. Evaluation of the PaO2/FiO2 ratio after cardiac surgery as a predictor of outcome during hospital stay[J]. BMC Anesthesiol, 2014, 14:83.
doi: 10.1186/1471-2253-14-83
[33] Hu BJ, Tao LL, Rosenthal VD, et al. Device-associated infection rates, device use, length of stay, and mortality in intensive care units of 4 Chinese hospitals: International Nosocomial Control Consortium findings[J]. Am J Infect Control, 2013, 41(4):301-306.
doi: 10.1016/j.ajic.2012.03.037
[34] 徐文秀, 方理超, 刘励军. APACHE Ⅱ评分和 SAPSⅡ评分预测危重症患者病死率的应用价值分析[J]. 中国血液流变学杂志, 2010, (2):245-247.
[35] 汪洋, 陈上仲, 胡才宝, 等. 基于随机森林法的严重脓毒症/脓毒性休克预后评估模型对患者28 d死亡的预测价值[J]. 中华危重病急救医学, 2017, 29(12):1071-1076.
[1] 许克新,丁泽华. 人工智能在功能泌尿外科的应用[J]. 北京大学学报（医学版）, 2023, 55(5): 771-774.
[2] 林瑜,吴静依,蔺轲,胡永华,孔桂兰. 基于集成学习模型预测重症患者再入重症监护病房的风险[J]. 北京大学学报（医学版）, 2021, 53(3): 566-572.
[3] 朱学华,杨明钰,夏海缀,何为,张智荧,刘余庆,肖春雷,马潞林,卢剑. 机器学习模型在预测肾结石输尿管软镜碎石术后早期结石清除率中的应用[J]. 北京大学学报（医学版）, 2019, 51(4): 653-659.
[4] 蔺轲，谢俊卿，胡永华，孔桂兰. 支持向量机在ICU急性肾损伤患者住院死亡风险预测中的应用[J]. 北京大学学报(医学版), 2018, 50(2): 239-244.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
ISSN 1671-167X
CN 11-4691/R
主管单位：中华人民共和国教育部
主办单位：北京大学
地址： 北京市海淀区学院路38号 北京大学医学部 《北京大学学报（医学版）》
邮编：100191
电话：010-82801551
Email： xbbjb2@bjmu.edu.cn

《北京大学学报（医学版）》
微信公众号
版权所有 © 2018 《北京大学学报（医学版）》编辑部