目的 基于血液常规指标联合系统性免疫炎症指数（SII）和系统性炎症反应指数（SIRI）构建较优的2型糖尿病（T2DM）患者发生糖尿病肾脏疾病（DKD）的风险预测模型并比较不同机器学习模型的预测性能，为DKD的早期筛查提供有效工具。 方法 回顾性选取2023年1月至2024年11月于广州医科大学附属中医医院住院的T2DM患者658例作为研究对象。采用计算机简单随机抽样法按7∶3的比例将患者数据分为训练集（460例）和测试集（198例）。首先，采用LASSO回归对包括SII、SIRI在内的12项潜在预测特征进行特征筛选；然后，基于筛选出的变量，分别应用逻辑回归（LR）、支持向量机（SVM）、随机森林（RF）和极限梯度提升（XGBoost）4种机器学习算法构建T2DM患者发生DKD的风险预测模型。使用AUC值、敏感度、准确度、F1分数等指标综合评价模型的区分度，并通过校准曲线和决策曲线分析（DCA）分别评估模型的校准度和临床实用性。最后，采用SHAP法对最优模型进行可解释性分析。 结果 经LASSO回归筛选出10项预测特征。SHAP值显示肌酐在4种风险预测模型中均具有较高的重要性；LR、RF、SVM和XGBoost4种模型在测试集的AUC值分别为0.914、0.943、0.929和0.917，F1分数分别为0.627、0.737、0.474和0.772。从混淆矩阵热力图中得到的总准确率分别为90.4%、92.4%、89.9%和93.4%；RF和XGBoost在混淆矩阵热力图中对DKD发生预测准确率较高分别为72.4%和75.9%。DCA显示4种模型在大部分阈值概率下均具有正的临床净收益。 结论 RF和XGBoost模型能较为准确地预测T2DM发生DKD的风险，有助于临床医生早期识别T2DM发生DKD的高风险患者。

Objective To construct an optimal risk prediction model for diabetic kidney disease （DKD） in patients with type 2 diabetes mellitus （T2DM） based on routine blood indicators combined with the systemic immune-inflammation index （SII） and systemic inflammation response index （SIRI）， and to compare the predictive performance of different machine learning models， so as to provide an effective tool for the early screening of DKD. Methods A total of 658 T2DM patients hospitalized in the Affiliated Hospital of Traditional Chinese Medicine of Guangzhou Medical University from January 2023 to November 2024 were retrospectively selected as the research subjects. The patient data were divided into a training set （460 cases） and a test set （198 cases） at a ratio of 7∶3 using the computer simple random sampling method. First， LASSO regression was used to screen 12 potential predictive features including SII and SIRI. Then， based on the screened variables， four machine learning algorithms， namely logistic regression （LR）， support vector machine （SVM）， random forest （RF）， and extreme gradient boosting （XGBoost）， were applied to construct risk prediction models for DKD in T2DM patients. Indicators such as AUC value， sensitivity， accuracy， and F1 score were used to comprehensively evaluate the discrimination of the models， and the calibration curve and decision curve analysis （DCA） were used to evaluate the calibration and clinical utility of the models respectively. Finally， the SHAP method was used to conduct interpretability analysis of the optimal model. Results Ten predictive features were screened out by LASSO regression. SHAP values showed that creatinine had high importance in all four risk prediction models. The AUC values of the LR， RF， SVM， and XGBoost models in the test set were 0.914， 0.943， 0.929， and 0.917 respectively， and the F1 scores were 0.627， 0.737， 0.474， and 0.772 respectively. The overall accuracies obtained from the confusion matrix heat maps were 90.4%， 92.4%， 89.9%， and 93.4% respectively. The prediction accuracies of RF and XGBoost for DKD occurrence in the confusion matrix heat maps were relatively high， at 72.4% and 75.9% respectively. DCA showed that all four models had positive clinical net benefits at most threshold probabilities. Conclusion The RF and XGBoost models can accurately predict the risk of DKD in T2DM patients， which is helpful for clinicians to identify high-risk T2DM patients with DKD at an early stage.