目的  基于机器学习(machine learning，ML)算法构建腹膜透析(peritoneal dialysis，PD)患者发生睡眠障碍的结局预测模型。 方法 采用多中心横断面调查研究，选择2024年1月─12月新疆维吾尔自治区人民医院、喀什地区第二人民医院等20家医院收治的578例PD患者为研究对象，根据匹兹堡睡眠质量指数(Pittsburgh sleep quality index，PSQI)，分为非睡眠障碍组(总分＜7分)与睡眠障碍组(≥7分)，按6:4随机分为训练集与内部验证集。采用单因素逻辑回归(univariate logistic regression，ULR)、基于随机森林算法的Boruta算法及最小绝对收缩和选择算子(least absolute shrinkage and selection operator，LASSO)识别关键预测指标，取并集后通过多因素逻辑回归(multivariate logistic regression，MLR)作进一步筛选，并将最终保留的特征变量用于训练6种机器学习模型，选出最佳模型并通过Shapley加法解释(shapley additive explanations，SHAP)评估其可解释性。 结果  MLR证实：年龄(OR=1.029，95% CI:1.011～1.049，P=0.002)、皮肤瘙痒(OR=2.266，95% CI:1.325～3.970，P=0.003)、焦虑(OR=4.075，95% CI: 2.399～7.012，P＜0.001)、尿素清除指数(OR=2.065，95% CI:1.210～3.577，P=0.008)和血钙(OR=7.476，95% CI:2.614～22.610，P＜0.001)是PD患者发生睡眠障碍的独立危险因素。在构建的模型中，极端梯度提升机(extreme gradient boosting，XGB)模型表现最佳，训练集的受试者工作特征(receiver operating characteristic，ROC)曲线下面积为0.830(95% CI:0.780～0.870)，内部验证集ROC曲线下面积为0.660(95% CI:0.560～0.760)，准确性为0.800。SHAP图显示，总蛋白、血钙和年龄是最重要的3个预测特征。 结论  本研究基于XGB算法构建了我国PD患者发生睡眠障碍的预测模型，有助于早期识别高危人群并支持精准干预。

Objective  To develop a prediction model for sleep disorders in maintenance peritoneal dialysis (PD) patients based on machine learning (ML) algorithms.  Methods  A multicenter cross-sectional study was conducted, enrolling 578 PD patients admitted to 20 hospitals, including Xinjiang Uygur Autonomous Region People's Hospital and The Second People's Hospital of Kashi Prefecture, from January to December 2024. Based on the pittsburgh sleep quality index (PSQI), the patients were divided into a non-sleep disorder group (total score ＜ 7) and a sleep disorder group (score≥ 7), and randomly assigned to a training set and an internal validation set at a 6:4 ratio. Univariate logistic regression (ULR), the random forest-based Boruta algorithm, and the least absolute shrinkage and selection operator (LASSO) were used to identify key predictive indicators. The union of these variables was further screened through multivariate logistic regression (MLR). The ultimately retained feature variables were used to train six machine learning models to identify the optimal model, and its interpretability was evaluated using Shapley additive explanations (SHAP). Results  MLR confirmed that age (OR=1.029, 95% CI: 1.011～1.049, P=0.002), pruritus (OR=2.266, 95%CI: 1.325～3.970, P=0.003), anxiety (OR=4.075, 95% CI: 2.399～7.012, P＜0.001), Kt/V (OR=2.065, 95%CI: 1.210～3.577, P=0.008), and serum calcium (OR=7.476, 95% CI: 2.614～22.610, P＜0.001) were independent risk factors for sleep disorders in PD patients. Among the constructed models, the extreme gradient boosting (XGB) model performed best, achieving an area under the receiver operating characteristic (ROC) curve of 0.830 (95% CI: 0.780～0.870) in the training set and 0.660 (95%CI: 0.560～0.760) in the internal validation set, with an accuracy of 0.800. SHAP plots indicated that total protein, serum calcium, and age were the three most important predictive features.   Conclusions  This study successfully developed the first XGB-based prediction model for sleep disorders among PD patients in China, facilitating the early identification of high-risk populations and supporting precise interventions.