添加链接
link之家
链接快照平台
  • 输入网页链接,自动生成快照
  • 标签化管理网页链接
The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely. As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health.
Learn more: PMC Disclaimer
Nan Fang Yi Ke Da Xue Xue Bao. 2023 Jul 20; 43(7): 1241–1247.
PMCID: PMC10366517

Language: Chinese | English

预测重症缺血性脑卒中死亡风险的模型:基于内在可解释性机器学习方法

An interpretable machine learning-based prediction model for risk of death for patients with ischemic stroke in intensive care unit

罗 枭

海军军医大学卫勤系军队卫生统计学教研室,上海 200433, Department of Military Health Statistics, Naval Medical University, Shanghai 200433, China

Find articles by 罗 枭

程 义

海军军医大学卫勤系军队卫生统计学教研室,上海 200433, Department of Military Health Statistics, Naval Medical University, Shanghai 200433, China

Find articles by 程 义

吴 骋

海军军医大学卫勤系军队卫生统计学教研室,上海 200433, Department of Military Health Statistics, Naval Medical University, Shanghai 200433, China

Find articles by 吴 骋

贺 佳

海军军医大学卫勤系军队卫生统计学教研室,上海 200433, Department of Military Health Statistics, Naval Medical University, Shanghai 200433, China 海军军医大学卫勤系军队卫生统计学教研室,上海 200433, Department of Military Health Statistics, Naval Medical University, Shanghai 200433, China COPD: Chronic obstructive pulmonary disease; WBC: White blood cell.Gender<0.001   Female784 (46.7%)376 (54.5%)   Male895 (53.3%)314 (45.5%)Age (year)69.5±14.378.8±11.6<0.001Weight (kg)81.9±21.172.9±18.9<0.001Smoking history0.9   Yes306 (18.2%)128 (18.6%)   No1373 (81.8%)562 (81.4%)Ventilation<0.001   Yes927 (55.2%)319 (46.2%)   No752 (44.8%)371(53.8%)Hypertension<0.001   Yes927 (55.2%)319(46.2%)   No752 (44.8%)371 (53.8%)Hyperlipidemia<0.001   Yes1148 (68.4%)410 (59.4%)   No531 (31.6%)280 (40.6%)Diabetes0.02   Yes402 (23.9%)198 (28.7%)   No1277 (76.1%)492 (71.3%)COPD0.02   Yes402 (23.9%)198 (28.7%)   No1277 (76.1%)492 (71.3%)Coronary heart disease0.34   Yes620 (36.9%)270 (39.1%)   No1059 (63.1%)420 (60.9%)Atrial fibrillation0.004   Yes528 (31.4%)312 (45.2%)   No1151 (68.6%)378 (54.8%)Heart rate (rate/min)77.9±14.085.2±15.8<0.001Diastolic blood pressure (mmHg) 68.7±13.166.9±12.10.002Systolic blood pressure (mmHg) 130.6±18.7130.2±20.30.67Respiratory rate (rate/min)18.5±3.020.2±3.8<0.001Temperature (℃)36.8±0.436.9±0.50.03Oxygen saturation (%)96.8±1.797.2±2.00.003Glucose (mg/dL)147.9±77.9174.3±96.1<0.001WBC (109/L)11.8±5.513.3±6.3<0.001Blood urea nitrogen (mg/dL)22.2±15.132.3±24.2<0.001Serum creatinine (mg/dL)1.2±1.21.5±1.2<0.001Sodium (mmol/L)140.2±3.6141.0±5.50.003International Normalized Ratio1.4±0.71.5±1.00.002Partial thromboplastin time (s) 36.9±19.436.8±17.30.9Erythrocyte specific volume (%) 34.2±7.033.1±6.4<0.001Platelets (109/L)204.3±79.5212.9±94.00.04Anion gap (mmol/L)13.1±2.914.3±3.3<0.001Bicarbonate (mmol/L)22.5±3.421.7±4.5<0.001Calcium (mmol/L)8.5±0.88.4±0.70.07Chloride (mmol/L)102.9±4.9102.1±5.90.004Potassium (mmol/L)3.9±0.53.9±0.60.27SOFA score3.4±2.65.5±3.4<0.001

2.2. 机器学习模型表现

将所有变量纳入机器学习模型中,经过超参数寻优选择4种方法各自表现最优的模型,4种模型应用于测试集数据( 表 2 ),其中EBM模型表现最好。AUC值从高到低依次排列为EBM(0.857)、RF(0.838)、LR(0.807)、Naive Bayes(0.785)( 图 2 ),经DeLong检验除EBM与朴素贝叶斯之间AUC值差异有统计学意义( P <0.05)外,其余模型的两两检验差异无统计学意义。Brier值从低到高依次排列为EBM(0.135)、RF(0.148)、LR(0.158)、Naive Bayes(0.200),从4种预测模型的校准图中可见EBM模型校准表现最好( 图 3 ),其斜率与截距最小,Naive Bayes表现最差。从4种预测模型的决策曲线分析结果中可见当概率阈值为0.10~0.80时,EBM的净获益率要高于其他模型( 图 4 )。

表 2

四种模型在测试集上的性能表现

Performance of the 4 models on the test set

Indicator LR Naive bayes RF EBM
LR: Logistic regression; RF: Random forest; EBM: Explainable boosting machine; AUC: Area under the subject's working characteristic curve.
AUC (95% CI ) 0.807 (0.773-0.836) 0.785 (0.755-0.813) 0.838 (0.810-0.870) 0.857 (0.831-0.887)
Accuracy 0.789 0.747 0.787 0.808
Precision 0.671 0.580 0.734 0.733
Recall 0.766 0.471 0.420 0.536
F1-score 0.488 0.520 0.535 0.619
Brier score 0.158 0.200 0.148 0.135
An external file that holds a picture, illustration, etc. Object name is nfykdxxb-43-7-1241-2.jpg

四种预测模型的受试者工作特征曲线

Receiver operating characteristic (ROC) curves of the 4 prediction models.

四种预测模型的校准曲线

Calibration curves of the 4 prediction models.

An external file that holds a picture, illustration, etc. Object name is nfykdxxb-43-7-1241-4.jpg

四种预测模型的决策曲线

Decision curve analysis for the 4 prediction models.

2.3. 变量重要性分析

EBM模型中各变量平均权重的绝对值越大对预后预测的影响越大( 图 5 )。排名前15的变量包括年龄(0.370),SOFA评分(0.284),平均心率(0.198),机械通气(0.197),平均呼吸频率(0.197),体质量(0.167),平均血氧饱和度(0.157),高血压(0.142),碳酸氢盐最小值(0.131),阴离子间隙最小值(0.127),钠离子最大值(0.125),血糖最大值(0.121),氯化物最小值(0.120),高脂血症(0.120)和平均舒张压(0.109)。其中最重要的前两个变量分别为年龄和SOFA评分,年龄和SOFA评分的风险效应趋势图中显示年龄越大,SOFA评分越高,模型预测得分越高,患者预后结局越差( 图 6 )。

EBM模型所选前15个变量的相对重要性得分

Relative importance scores of the top 15 variables selected by the EBM model.

风险效应趋势图

Risk effect trend graph. A : Trend graph of risk effect of age. B : Trend graph of risk effect of SOFA score.

3. 讨论

本研究利用MIMIC-Ⅳ数据库中重症缺血性脑卒中患者数据构建了一个EBM模型,以预测缺血性脑卒中一年的死亡风险。基于公开数据库,建模所涉及的信息是客观的,样本量充足,且容易获得。本研究构建的EBM模型具有良好的区分度(AUC=0.857)与校准度(Brier score=0.135),且具有较强的可解释性,综合表现优秀,可辅助医生进行决策,达到了预期目标。

重症缺血性脑卒中是一种多因素疾病,若干危险因素对其结果有影响。根据EBM模型变量重要性分析结果显示:年龄,SOFA评分,平均心率,机械通气,平均呼吸频率等因素与重症缺血性中风患者生存率密切相关。一般来说,脑卒中是一种衰老性疾病。此外,衰老会加重脑卒中后的脑损伤,与衰老相关的脑内基础Bcl-2基因表达的减少会增加脑卒中后的细胞凋亡从而加重神经损伤 [ 25 ] 。相关文献报道年龄是缺血性脑卒中独立的危险因素。随着年龄的增长,大脑储备能力减弱,导致老年患者的预后恶化 [ 26 ] 。一项关于继发性腹膜炎重症患者死亡风险因素的研究显示 [ 27 ] ,随着SOFA评分的增加,提示感染加剧,患者死亡率呈阶梯式上升。另有研究显示 [ 28 ] ,在入院24 h内SOFA评分增加2分或以上对院内死亡率及预测ICU住院时间的综合结果相比系统性炎症反应综合征标准(SIRS)及qSOFA评分更好。根据Etienne等 [ 29 ] 的一项多中心队列研究,在需要机械通气的急性脑卒中患者中,插管的原因和是否接受急性期插管治疗与其一年生存率独立相关。其余预测指标也与既往研究相似 [ 5 , 30 ]

近年来,使用机器学习方法预测脑卒中预后的研究不断涌现。Thomas等 [ 18 ] 使用奥地利脑卒中登记中心的数据预测缺血性卒中患者入院早期死亡风险,在这项研究中正则化Logistic回归预测缺血性脑卒中患者早期死亡的AUC可达0.88;Fernandez等 [ 19 ] 使用随机森林模型预测脑卒中三个月内的死亡率和功能结局,模型具有良好的预测能力,AUC最高为0.91;Lehmann等 [ 31 ] 使用机器学习方法构建临床、影像学和生物标志物联合预测缺血性脑卒中1年内死亡模型,其中神经网络模型达到最高准确性,AUC为0.97;有研究基于前瞻性队列研究使用机器学习预测3个月后Rankin量表评分结果,结果显示深度神经网络模型的AUC可达0.88 [ 32 ] 。以上研究除使用本研究中的变量外,还包括脑卒中的特定变量,如NIHSS评分或神经影像学结果等,这可能是导致模型预测效能高于本研究的原因之一。但模型的准确性指标只是模型预测能力的部分体现,如果模型缺乏可解释性也很难应用于临床实践,本研究使用内在可解释性模型在达到与复杂黑箱模型相似准确性的前提下,增加了模型的可解释性,有利于增强临床使用者对预测结果的信心与信任。

尽管本研究模型表现良好,但也有其局限性:(1)受限于数据库信息本研究未纳入与缺血性脑卒中相关的NIHSS评分、神经影像学数据等;(2)本研究基于MIMIC-Ⅳ公共数据库,尚未使用我国数据对模型进行外部验证;(3)尚未使用最近提出的其他内在可解释模型例如基于规则的Falling Rule Lists模型、基于神经网络的GAMI-Net模型等进行对比验证。

综上所述,本研究基于MIMIC-Ⅳ数据库构建了一个用于预测重症缺血性脑卒中患者一年死亡率的内在可解释机器学习模型,该模型分类性能良好、表现稳定,且具有较强的可解释性,可以辅助医生对患者进行合理的预后评估与治疗。在今后的研究中考虑纳入临床轨迹、神经影像学信息及用药信息等提升变量数据维度,同时也需要使用更具代表性的外部数据来进一步支持我们的结论。

Biography

罗枭,在读硕士研究生,E-mail: moc.361@105039oaixoul

Funding Statement

上海市公共卫生体系建设学科带头人计划(GWV-10.2-XD05);上海市产业协同创新项目(2021-cyxt1-kj10)

References

1. 《中国脑卒中防治报告》编写组 《中国脑卒中防治报告2019》概要 中国脑血管病杂志 2020; 17 (5):271–81. [ Google Scholar ]
2. 彭 斌, 刘 鸣, 崔 丽英. 与时俱进的新指南: 《中国急性缺血性脑卒中诊治指南2018》解读 中华神经科杂志 2018; 51 (9):666–82. doi: 10.3760/cma.j.issn.1006-7876.2018.09.004. [ CrossRef ] [ Google Scholar ]
3. 赵 楠楠, 郑 印, 黄 穹琼, et al. 急性缺血性脑卒中早期预后不良的危险因素分析及预测模型构建 中国临床研究 2022; 35 (4):456–461. doi: 10.13429/j.cnki.cjcr.2022.04.003. [ CrossRef ] [ Google Scholar ]
4. Chang HY, Jung CK, Woo JI, et al. Artificial intelligence in pathology. J Pathol Transl Med. 2019; 53 (1):1–12. doi: 10.4132/jptm.2018.12.16. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
5. Vi da, Abedi, P hD, et al. Predicting short and long-term mortality after acute ischemic stroke using EHR. J Neurol Sci. 2021; 427 :117560. doi: 10.1016/j.jns.2021.117560. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
6. Kurtz P, Peres IT, Soares M, et al. Hospital length of stay and 30-day mortality prediction in stroke: a machine learning analysis of 17, 000 ICU admissions in Brazil. Neurocrit Care. 2022; 37 (2):313–21. [ PubMed ] [ Google Scholar ]
7. Reddy S. Explainability and artificial intelligence in medicine. Lancet Digit Health. 2022; 4 (4):e214–e215. doi: 10.1016/S2589-7500(22)00029-2. [ PubMed ] [ CrossRef ] [ Google Scholar ]
8. Ibrahim M, Louie M, Modarres C, et al. Global explanations of neural networks: mapping the landscape of predictions[C]//Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society. January 27-28, 2019, Honolulu, HI, USA. New York: ACM, 2019: 279-87.
9. Stinear CM, Smith MC, Byblow WD. Prediction tools for stroke rehabilitation. Stroke. 2019; 50 (11):3314–22. doi: 10.1161/STROKEAHA.119.025696. [ PubMed ] [ CrossRef ] [ Google Scholar ]
10. Miller T. Explanation in artificial intelligence: insights from the social sciences. Artif Intell. 2019; 267 :1–38. doi: 10.1016/j.artint.2018.07.007. [ CrossRef ] [ Google Scholar ]
11. Johnson A, Bulgarelli L, Pollard T, et al. MIMIC-Ⅳ (version 2.0) [Z]. PhysioNet. 2022
12. Moons KG, Altman DG, Reitsma JB, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med. 2015; 162 (1):W1–W73. doi: 10.7326/M14-0698. [ PubMed ] [ CrossRef ] [ Google Scholar ]
13. Deshmukh F, Merchant SS. Explainable machine learning model for predicting GI bleed mortality in the intensive care unit. Am J Gastroenterol. 2020; 115 (10):1657–68. doi: 10.14309/ajg.0000000000000632. [ PubMed ] [ CrossRef ] [ Google Scholar ]
14. Miettinen T, Mäntyselkä P, Hagelberg N, et al. Machine learning suggests sleep as a core factor in chronic pain. Pain. 2021; 162 (1):109–23. doi: 10.1097/j.pain.0000000000002002. [ PubMed ] [ CrossRef ] [ Google Scholar ]
15. Walfish S. A review of statistical outlier methods. Pharmaceutical technology. 2006; 30 (11):82. [ Google Scholar ]
16. Nori H, Jenkins S, Koch P, et al. Interpretml: A unified framework for machine learning interpretability. arXiv preprint arXiv. 1909; 09223 :2019. [ Google Scholar ]
17. Luo W, Phung D, Tran T, et al. Guidelines for developing and reporting machine learning predictive models in biomedical research: a multidisciplinary view. J Med Internet Res. 2016; 18 (12) [ PMC free article ] [ PubMed ] [ Google Scholar ]
18. Gattringer T, Posekany A, Niederkorn K, et al. Predicting early mortality of acute ischemic stroke: score-based approach. Stroke. 2019; 50 (2):349–356. doi: 10.1161/STROKEAHA.118.022863. [ PubMed ] [ CrossRef ] [ Google Scholar ]
19. Fernandez-Lozano C, Hervella P, Mato-Abad V, et al. Random forestbased prediction of stroke outcome. Sci Rep. 2021; 11 (1):10071. doi: 10.1038/s41598-021-89434-7. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
20. Qu YJ, Deng XL, Lin S, et al. Using innovative machine learning methods to screen and identify predictors of congenital heart diseases. Front Cardiovasc Med. 2021; 8 :797002. [ PMC free article ] [ PubMed ] [ Google Scholar ]
21. DeLong E, DeLong D, Clarke‐Pearson D. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988:837–45. [ PubMed ] [ Google Scholar ]
22. Powers D M W. Evaluation: from precision, recall and F-measure to ROC, informedness, markedness and correlation. arXiv preprint arXiv. 2010; 16061 :2020. [ Google Scholar ]
23. Ralston A, Liu P, Warrener K, et al. Small field diode correction factors derived using an air core fibre optic scintillation dosimeter and EBT2 film. Phys Med Biol. 2012; 57 (9):2587–602. doi: 10.1088/0031-9155/57/9/2587. [ PubMed ] [ CrossRef ] [ Google Scholar ]
24. van Calster B, Wynants L, Verbeek JFM, et al. Reporting and interpreting decision curve analysis: a guide for investigators. Eur Urol. 2018; 74 (6):796–804. doi: 10.1016/j.eururo.2018.08.038. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
25. P, Do ng. Aging causes exacerbated ischemic brain injury and failure of sevoflurane post-conditioning: role of B-cell lymphoma-2. Neuroscience. 2014; 275 :2–11. doi: 10.1016/j.neuroscience.2014.05.064. [ PubMed ] [ CrossRef ] [ Google Scholar ]
26. Luedi R, Hsieh K, Slezak A, et al. Age dependency of safety and outcome of endovascular therapy for acute stroke. J Neurol. 2014; 261 (8):1622–7. doi: 10.1007/s00415-014-7401-0. [ PubMed ] [ CrossRef ] [ Google Scholar ]
27. De Pascale G, Antonelli M, Deschepper M, et al. Poor timing and failure of source control are risk factors for mortality in critically ill patients with secondary peritonitis. Intensive Care Med. 2022; 48 (11):1593–606. doi: 10.1007/s00134-022-06883-y. [ PubMed ] [ CrossRef ] [ Google Scholar ]
28. Raith EP, Udy AA, Bailey M, et al. Prognostic accuracy of the SOFA score, SIRS criteria, and qSOFA score for In-hospital mortality among adults with suspected infection admitted to the intensive care unit. JAMA. 2017; 317 (3):290–300. doi: 10.1001/jama.2016.20328. [ PubMed ] [ CrossRef ] [ Google Scholar ]
29. de Montmollin E, Terzi N, Dupuis C, et al. One-year survival in acute stroke patients requiring mechanical ventilation: a multicenter cohort study. Ann Intensive Care. 2020; 10 (1):53. doi: 10.1186/s13613-020-00669-5. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
30. Park D, Jeong E, Kim H, et al. Machine learning-based three-month outcome prediction in acute ischemic stroke: a single cerebrovascular-specialty hospital study in South Korea. Diagnostics (Basel) 2021; 11 (10):1909. doi: 10.3390/diagnostics11101909. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
31. Lehmann ALCF, Alfieri DF, de Araújo MCM, et al. Immuneinflammatory, coagulation, adhesion, and imaging biomarkers combined in machine learning models improve the prediction of death 1 year after ischemic stroke. Clin Exp Med. 2022; 22 (1):111–23. doi: 10.1007/s10238-021-00732-w. [ PubMed ] [ CrossRef ] [ Google Scholar ]
32. Heo JN, Yoon JG, Park H, et al. Machine learning-based model for prediction of outcomes in acute stroke. Stroke. 2019; 50 (5):1263–5. doi: 10.1161/STROKEAHA.118.024293. [ PubMed ] [ CrossRef ] [ Google Scholar ]

Articles from Journal of Southern Medical University are provided here courtesy of Editorial Department of Journal of Southern Medical University