摘要目的:建立急性缺血性卒中（acute ischemic stroke, AIS）患者发病后3个月卒中后抑郁（post-stroke depression, PSD）的机器学习预测模型。方法:回顾性纳入2021年1月至2023年12月期间合肥市第二人民医院收治的AIS患者。根据发病后3个月17项汉密尔顿抑郁量表（Hamilton Depression Rating Scale, HAMD）评价结果分为PSD组和非PSD组。采用递归特征消除（recursive feature elimination, RFE）方法筛选PSD的特征变量。基于 logistic回归（ logistic regression, LR）、随机森林（random forest, RF）和支持向量机（supported vector machine, SVM）3种机器学习算法构建AIS患者PSD预测模型，并通过受试者工作特征（receiver operating characteristic, ROC）曲线下面积（area under curve, AUC）以及校准曲线评估模型性能。采用沙普利加和解释（SHapley Additive exPlanations, SHAP）算法分析各危险因素的贡献度。 结果:共纳入243例AIS患者，男性159例（64.6%），年龄（64.32±11.54）岁，中位受教育年限为6年，13例（5.3%）独居；105例（42.7%）有既往卒中史；中位基线美国国立卫生研究院卒中量表（National Institutes of Health Stroke Scale, NIHSS）评分3分，中位基线改良Rankin量表（modified Rankin Scale, mRS）2分；33例（13.4%）接受了静脉溶栓治疗。93例（38.27%）在发病后3个月时存在PSD。RFE显示，最佳特征数量为11个，分别是基线NIHSS评分、基线mRS评分、C反应蛋白、静脉溶栓、低密度脂蛋白胆固醇、小血管闭塞、D-二聚体、总胆固醇、饮酒、右侧梗死和基线收缩压。ROC曲线分析显示，RF模型的预测效果最佳（AUC=0.831，95%置信区间0.730~0.931），其次是SVM模型（AUC=0.827，95%置信区间0.713~0.941），LR模型预测性能最低（AUC=0.771，95%置信区间0.658~0.885）。校准曲线显示，RF模型与理想曲线贴切良好，因此为最终优势模型。SHAP显示，基线NIHSS评分、基线mRS评分、低密度脂蛋白胆固醇、总胆固醇和静脉溶栓的贡献度排在前5位。结论:RF模型可有效预测PSD的发生风险，基线NIHSS评分、基线mRS评分、低密度脂蛋白胆固醇和总胆固醇以及静脉溶栓是关键预测因素。

abstractsObjectives:To develop a machine learning prediction model for post-stroke depression (PSD) in patients with acute ischemic stroke (AIS) at 3 months after onset.Methods:Patients with AIS admitted to the Second People's Hospital of Hefei from January 2021 to December 2023 were included retrospectively. According to the 17-item Hamilton Depression Rating Scale (HAMD) evaluation results at 3 months after onset, they were divided into PSD group and non-PSD group. The recursive feature elimination (RFE) method was used to screen the characteristic variables of PSD. A PSD prediction model for patients with AIS was developed based on three machine learning algorithms: logistic regression (LR), random forest (RF), and supported vector machine (SVM). The area under a receiver operating characteristic (ROC) curve (AUC) and calibration curve were used to evaluate the performance of the model. The SHapley Additive exPlanations (SHAP) algorithm was used to analyze the contribution of each risk factor. Results:A total of 243 patients with AIS were included, including 159 males (64.6%), aged 64.32±11.54 years, the median years of schooling was 6 years, and 13 males (5.3%) lived alone. 105 patients (42.7%) had a history of stroke. The median baseline National Institutes of Health Stroke Scale (NIHSS) score was 3, and the median baseline Modified Rankin Scale (mRS) score was 2. 33 patients (13.4%) received intravenous thrombolysis treatment. 93 patients (38.27%) had PSD at 3 months after onset. RFE showed that the optimal number of features was 11, including baseline NIHSS score, baseline mRS score, C-reactive protein, intravenous thrombolysis, low-density lipoprotein cholesterol, small vessel occlusion, D-dimer, total cholesterol, alcohol consumption, right side infarction, and baseline systolic blood pressure. ROC curve analysis shows that the RF model had the best predictive performance (AUC=0.831, 95% confidence interval 0.730-0.931), followed by the SVM model (AUC=0.827, 95% confidence interval 0.713-0.941), and the LR model has the lowest predictive performance (AUC=0.771, 95% confidence interval 0.658-0.885). The calibration curve shows that the RF model fits well with the ideal curve, making it the final advantageous model. SHAP showed that the contribution of baseline NIHSS score, baseline mRS score, low-density lipoprotein cholesterol, total cholesterol, and intravenous thrombolysis ranked among the top 5.Conclusions:The RF model can effectively predict the risk of PSD. The baseline NIHSS score, baseline mRS score, low-density lipoprotein cholesterol, and total cholesterol, as well as intravenous thrombolysis are the key predictive factors.