【作者】 冯峰；

【导师】 杨向东；

【作者基本信息】 山东大学 ， 内科学·肾脏病（专业学位）， 2022， 硕士

【摘要】 研究背景和目的特发性膜性肾病(idiopathic membranous nephropathy,IMN)是目前成年人发生肾病综合征的主要病因之一,其病程变化及预后较为多变。尿白蛋白肌酐比(urine albumin-to-creatinine ratio,UACR)是评价IMN严重程度的主要指标之一,也是临床选择治疗方案的重要参考。因此,本研究通过对IMN患者的回顾性分析,利用机器学习(machine learning,ML)预测患者早期UACR并分析影响预后的相关因素,有助于提示临床医生筛查预后不良和治疗效果欠佳患者,为加强此类患者管理及及时调整治疗方案提供临床依据。研究方法通过检索并筛选出2013年3月至2021年4月于山东大学齐鲁医院初次就诊并在院内行肾穿刺活检病理确诊为IMN,且其随访月份的预测指标完整的患者,总计患者383例。收集并记录电子病历中患者的人口学信息、肾脏活检病理结果和各项检验数据。将随机筛选2013年3月至2021年4月入组的300例IMN患者的数据作为训练集,将随机筛选2013年3月至2021年4月纳入的83例IMN患者的数据作为测试集。本研究采用了六种最新的机器学习算法和并由此建立一种混合算法,将测试集中IMN患者基线各项指标输入模型,预测其治疗后第1、3、6月的UACR,并将预测值与真实值进行对比以评估模型的准确性与稳定性。研究结果1.本研究共纳入383例IMN患者,通过机器学习算法可预测IMN患者1、3、6个月UACR。其中混合模型在预测中表现最好。基于基线数据的UACR预测相对于第1、3、6个月真实值的平均绝对误差(mean absolute errors,MAEs)分别为1.75(95%CI:1.53,1.97)、1.71(1.48,1.91)和 1.31(1.05,1.59);第 1 个月、3 个月和 6 个月 UACR 预报相对于地面真实值的 MSEs 分别为 4.74(4.52,4.96)、6.06(5.79,6.33)和 4.79(4.54,5.04)。2.通过机器学习发现基线水平UACR、基线血清免疫球蛋白G水平、甲状腺功能等在IMN早期预后预测中权重较大,提示我们以上指标可能是影响预后的重要因素。研究结论本研究利用患者基线临床病例资料和检验数据建立了一个准确的UACR机器学习预测模型,并验证了其对IMN患者UACR的个体预测准确性,并分析了影响预后的相关因素,该模型可协助指导临床医生筛选预后不良的患者,有助于临床医生及时调整和优化此类患者的治疗方案。更多还原

【Abstract】 Background and ObjectiveIdiopathic membranous nephropathy(IMN)is one of the main causes of nephrotic syndrome in adults.The course of IMN is quite variable.Urine albumin-to-creatinine ratio(UACR)is a main indicator to evaluate the severity of IMN and it is also an important reference for drug regimen.So,this study,a single center retrospective analysis of IMN patients,aimed to predict the UACR and analyze the risk factors of prognosis,which could potentially help to evaluate the progression and therapeutic effect in newly diagnosed IMN patients.MethodsThis study involved 383 IMN patients diagnosed by renal biopsy in Qilu Hospital of Shandong University.The real-world patients’ demographics,conclusion of renal biopsy and laboratory data were collected retrospectively from electronic medical record(EMR).The data of 300 IMN randomly selected patients from March 2013 to April 2021 were used as the training set.The data of 83 selected IMN patients from January March 2013 to April 2021 were regarded as the test set.Six different ML algorithms and a stack algorithm were used to predict post-therapeutic UACR in patients with IMN.The UACR predicted by ML algorithms was compared with the ground truth.ResultThe ML algorithms were able to accurately predict UACR of IMN patients 6 months in advance.The stack algorithm(Ridge and Lasso)performed best in UACR predictions.The mean absolute errors(MAEs)of UACR predictions based on the baseline data with respect to the ground truth were 1.75(95%CI:1.53,1.97),1.71(1.48,1.91)and 1.31(1.05,1.59)for the 1-,3-and 6-month predictions,respectively;and the MSEs of the UACR predictions with respect to the ground truth were 4.74(4.52,4.96),6.06(5.79,6.33)and 4.79(4.54,5.04)for the 1-,3-and 6-month predictions,respectively.Baseline UACR,baseline serum immunoglobulin G level,and thyroid function were found to be significant predictors of early prognosis of IMN by machine learning.ConclusionIn this retrospective study,a ML model was developed and validated for individual prediction of UACR in IMN patients.The ML model could be used in clinical practice to guide clinicians to screen patients with poor prognosis and optimize treatment readily.更多还原