摘要
目的探讨可解释性机器学习在预测脑卒中患者下肢深静脉血栓形成(DVT)风险的价值,为临床预防DVT提供参考。方法回顾性收集2021年1月至2025年4月于安徽中医药大学第二附属医院就诊的450例脑卒中患者的临床数据,经单因素分析和LASSO回归筛选变量后,将数据按8∶2拆分为训练集和验证集,并运用决策树、极限梯度提升算法、支持向量机、轻量级梯度提升机算法(LightGBM)、K-近邻算法5种机器学习方法在训练集构建模型,通过5折交叉验证对模型参数调优。在验证集上,曲线下面积、准确度、灵敏度等指标评估各模型性能,选取最优模型。最后运用Shapley加性解释(SHAP)算法对最优模型进行可解释性分析。结果本研究中DVT的发生率为13.56%。通过单因素分析和LASSO回归筛选出9个变量:脑卒中类型、脱水剂、苔色、年龄、日常生活能力量表(ADL)评分、血红蛋白、血小板计数、血肌酐、D-二聚体。在5种机器学习模型中表现最优为LightGBM模型(验证集AUC:0.903,准确度:0.857,精确度:0.971,F1分数:0.912)。SHAP算法分析结果显示,对预测脑卒中患者发生下肢DVT的贡献度排名前5位的特征为D-二聚体、ADL评分、苔色、血肌酐、血红蛋白。结论使用LightGBM构建的预测模型表现较佳,有助于临床工作者识别高危DVT风险的患者;D-二聚体、ADL评分等指标对预测脑卒中患者发生DVT有重要指导意义,可为DVT的预防提供一定的参考。
Objective To investigate the potential of interpretable machine learning in assessing the risk of lower extremity deep vein thrombosis(DVT)in stroke patients and to offer a reference for the clinical prevention of DVT.Methods The clinical data of 450 stroke patients at the Second Affiliated Hospital of Anhui University of Traditional Chinese Medicine from January 2021 to April 2025 were retrospectively analyzed.Following variable screening using univariate analysis and LASSO regression,the dataset was partitioned into a training set and a validation set in an 8∶2 ratio.Five machine learning algorithms—decision tree,extreme gradient boosting,support vector machine,light gradient boosting machine(LightGBM),and K-nearest neighbor classification were employed to construct predictive models on the training set.Model parameters were optimized via 5-fold cross-validation.On the validation set,the performance of each model was evaluated by indicators such as the area under the curve,accuracy,and sensitivity,and the optimal model was selected.Finally,the interpretability of the selected model was evaluated using the Shapley additive explanation(SHAP)algorithm.Results The incidence of DVT in this study was 13.56%.Univariate analysis and LASSO regression identified nine variables as significant predictors:stroke type,dehydrating agent,coating color,age,activities of daily living(ADL)score,hemoglobin level,platelet count,serum creatinine,and D-dimer.Among the five machine learning models evaluated,the LightGBM model demonstrated the best performance with an AUC of 0.903,accuracy of 0.857,precision of 0.971,and F1 score of 0.912.Furthermore,SHAP algorithm analysis revealed that the top five features contributing to the prediction of lower extremity DVT in stroke patients were D-dimer,ADL score,coating color,serum creatinine,and hemoglobin level.Conclusion The prediction model constructed using LightGBM has a good performance,which is helpful for clinicians to identify patients at high risk of DVT.D-dimer,ADL score and other indicators have important guiding significance in predicting the occurrence of DVT in stroke patients,which can provide certain reference for the prevention of DVT.
作者
夏爱芳
孙善斌
陈冲
江颖子
王婷
吴炳坤
李春标
梁月光
XIA Aifang;SUN Shanbin;CHEN Chong;JIANG Yingzi;WANG Ting;WU Bingkun;LI Chunbiao;LIANG Yueguang(Department of Rehabilitation,the Second Affiliated Hospital of Anhui University of Chinese Medicine,Anhui Province,Hefei 230001,China;Department of Encephalopathy,the Second Affiliated Hospital of Anhui University of Chinese Medicine,Anhui Province,Hefei 230001,China;Department of Nursing,the Second Affiliated Hospital of Anhui University of Chinese Medicine,Anhui Province,Hefei 230001,China;Department of Geriatrics,the Second Affiliated Hospital of Anhui University of Chinese Medicine,Anhui Province,Hefei 230001,China;School of Nursing,Anhui University of Chinese Medicine,Anhui Province,Hefei 230012,China)
出处
《中国医药导报》
2025年第24期34-40,共7页
China Medical Herald
基金
国家优势专科康复科建设项目(国中医药医政函〔2024〕90号)
安徽省省级临床重点专科建设项目(皖卫医秘〔2022〕105号)
安徽省高等学校科学研究项目(2023AH 050804)。
