摘要
目的利用套索(least absolute shrinkage and selection operator,LASSO)回归和贝叶斯网络分析方法,构建适合肿瘤患者急性肾损伤(acute kidney injury,AKI)发病风险的贝叶斯网络预测模型,为早期识别高危人群,制定AKI精准预防策略提供科学依据。方法以2014年10月1日至2015年9月30日在复旦大学附属中山医院就诊的恶性肿瘤住院患者为研究对象。于医院数据管理平台收集患者年龄、性别、体重指数、既往病史、肿瘤类型/治疗、基础肝肾功能、生化和电解质指标等数据资料。通过LASSO回归筛选出与AKI发生显著相关的影响因素;借助贝叶斯网络分析进一步描述变量间相互作用并评价模型预测效能。结果26914名研究对象中,AKI发病率为12.4%(n=3326),其中肾癌(27.3%),多发性骨髓瘤(24.1%)和急性粒细胞白血病(23.9%)患者的AKI发病率最高。LASSO回归筛选出22个与AKI发生相关性最显著的变量,包括年龄、性别、体重指数、糖尿病史、肿瘤类型/分期/治疗方式、肝功能、肾小球滤过率(estimated glomerular filtration rate,e GFR)/血清肌酐值/血尿酸、白蛋白、血红蛋白和白细胞计数、血钠/血钾等电解质。贝叶斯网络模型发现血红蛋白、e GFR、血氯和血磷与AKI的发生有直接联系;节点治疗方式通过影响血钠和白蛋白等中间节点间接影响AKI的发生;糖尿病和性别通过节点尿酸间接相连e GFR,而后者是AKI的父节点。模型推理在其他条件一致的情况下,贫血和e GFR≤59 m L·min-1·1.73 m-2的患者发生AKI的概率最高(55.7%);而上述指标均正常者AKI发病率最低(3.0%)。模型评价发现贝叶斯网络模型的分类准确率为88.8%,接受者操作特性曲线曲线下面积为0.806。结论基于LASSO变量选择联合贝叶斯网络分析构建的模型在肿瘤相关AKI的影响因素分析中更符合实际理论,其在发病风险预测中有较好的临床应用价值。
Objective To explore the associated risk factors of acute kidney injury(AKI)in cancer patients by least absolute shrinkage and selection operator(LASSO)regression-based Bayesian networks(BN),to estimate the prediction ability of BN model,and then to identify the high-risk patients of AKI through BN reasoning.Methods During Oct 1 st2014 and Sept 30 th2015,patients with malignancies were recruited in Zhongshan Hospital,Shanghai,China.Data on demographics,comorbidities,and clinicalrecords were exported from the hospital inpatient database.The candidate features for AKI was selected by LASSO regression and presented their inter-relationships in BN.Results Of 26914 eligible patients,3326 AKI cases were identified.The highest rates were localized to renal cancer(27.3%),multiple myeloma(24.1%),and leukemia(23.9%).The LASSO regression screened 22 candidate variables for further analysis,including age,gender,BMI,diabetes,cancer category/stage/treatment,liver dysfunction,estimated glomerular filtration rate(e GFR),serum creatinine(SCr),serum uric acid(SUA),albumin,hemoglobin and other biochemical indicators and electrolyte disorders.BN model revealed complex correlations between these related factors,in which a direct connection between hemoglobin,e GFR,serum chlorine&phosphorus and AKI were found.Treatment was indirectly linked to AKI through albumin and serum sodium,while diabetes and gender diabetes created connections with AKI through affecting SUA levels.Inferences by BN found that when poor e GFR and anemia probability occurred simultaneously,the probability of AKI may reach 55.7%.However,once these indicators were at the normal level,the estimate can be reduced to 3.0%.BN’s area under the ROC curve is 0.806,shared high classification accuracy(88.8%)to reflect the dependence among nodes.Conclusion Bayesian networks combined with LASSO can analyze not only how the correlative factors affect AKI but also their interrelationships.This model also shows good clinical application value in the prediction of AKI.
作者
李阳
陈晓泓
王一梅
胡家昌
沈子妍
沈波
林静
丁小强
LI Yang;CHEN Xiao-hong;WANG Yi-mei;HU Jia-chang;SHEN Zi-yan;SHEN Bo;LIN Jing;DING Xiao-qiang(Department of Nephrology,Zhongshan Hospital,Fudan University,Shanghai 200032,China)
出处
《复旦学报（医学版）》
CAS
CSCD
北大核心
2020年第4期521-530,共10页
Fudan University Journal of Medical Sciences
基金
复旦大学附属中山医院院级青年基金(2019ZSQN19)
上海市肾脏疾病与血液净化重点实验室科研项目(14DZ2260200)
上海市肾脏疾病临床医学中心科研项目(2017ZZ01015)。
