基于监督学习算法建立糖皮质激素联合环磷酰胺治疗特发性膜性肾病效果预测模型

Establishment of A Prediction Model for the Efficacy of Glucocorticoid Combined with Cyclophosphamide in the Treatment of Idiopathic Membranous Nephropathy Based on Supervised Learning Algorithm

目的建立基于监督学习算法的糖皮质激素(GC)联合环磷酰胺治疗特发性膜性肾病(IMN)效果的预测模型。方法入选2014年7月1日至2023年6月30日确诊的IMN患者,同时接受≥6个月GC联合环磷酰胺治疗,采集相关临床资料。运用Python软件构建9种监督学习模型,采用受试者工作特征曲线下面积(AUC)评估各模型的预测性能,筛选与疗效相关的指标,并根据结果构建预测工具。结果共纳入122例患者,其中57例(46.7%)完全缓解、39例(32.0%)部分缓解、26例(21.3%)未缓解。在纳入全部136项临床指标时,轻量级梯度提升机(LGBM)在9种监督学习模型中的AUC最高(0.965)。特征筛选结果显示第3个月的24 h尿蛋白定量(24 h UTP)下降率和血清白蛋白上升率与疗效的相关性最强。在仅纳入上述2个特征再次建模后,仍以LGBM的AUC最高(0.978)。故最终以LGBM为基础构建在线预测工具,网址为www.imnpredict.online。结论基于监督学习算法的GC联合环磷酰胺治疗IMN效果预测模型提示,治疗开始后第3个月的24 h UTP和血清白蛋白变化率是预测患者疗效的主要因素。该模型和在线工具可在IMN治疗早期对疗效进行预测,为患者个体化治疗提供参考。

Objective To establish a supervised learning algorithm-based prediction model for the efficacy of Glucocorticoid(GC)combined with Cyclophosphamide in the treatment of idiopathic membranous nephropathy(IMN).Methods Patients diagnosed with IMN from July 1,2014 to June 30,2023 were selected and treated with GC combined with Cyclophosphamide for≥6 months,and relevant clinical data were collected.Nine supervised learning models were constructed using Python software,and the predictive performance of each model was evaluated by using the area under the receiver operating characteristic(ROC)curve(AUC).Indicators related to efficacy were screened,and prediction tools were constructed according to the results.Results A total of 122 patients were included,of which 57(46.7%)had a complete response,39(32.0%)had a partial response,and 26(21.3%)had no response.When all 136 clinical measures were included,lightweight gradient boosting machine(LGBM)had the highest AUC(0.965)among the nine supervised learning models.The results of characteristic screening showed that the decrease rate of 24 h urinary protein quantification(24 h UTP)and the increase rate of serum albumin at 3 months after initiation of treatment had the strongest correlation with the efficacy.After re-modeling with only the above two features included,the AUC of LGBM was still the highest(0.978).Therefore,this study finally constructed an online prediction tool based on LGBM,and the website is www.imnpredict.online.Conclusion The prediction model of the efficacy of GC combined with Cyclophosphamide on IMN based on supervised learning algorithm suggests that 24 h UTP and serum albumin change rate at 3 months after the initiation of treatment are the main factors to predict the efficacy in the patients.The model and online tool can predict the efficacy in the treatment of early IMN and provide a reference for individualized treatment of patients.