基于机器学习的卒中后抑郁影响因素分析

Analysis of the Influencing Factors of Post-stroke Depression: Based on Machine Learning

目的通过机器学习判断脑卒中患者发生抑郁的影响因素。方法从病历系统中提取符合纳入条件的688例脑卒中患者的病历资料,包括年龄、性别、脉象、面色、舌质、舌苔、中医药干预方式、体重指数(BMI)、血压、血糖、血甘油三酯、血总胆固醇、吸烟史、饮酒史、抑郁家族史、影像学卒中病灶部位及最终抑郁判断结果。采用单规则(1R)算法进行学习并判断提取信息中影响卒中后患者是否抑郁的危险因素,继而将所搜集病例分为训练数据集(500例)和测试数据集(188例),并使用随机森林模型形成最优判别结果。结果通过单规则算法得出脑卒中后是否抑郁最重要的影响因素为卒中病灶所在部位,其中计算机推测卒中病灶位于额叶及颞叶者最易发生卒中后抑郁,基底节、脑干、小脑、延髓、枕叶的病灶则不易引起抑郁,其准确分类率达到88.95%(612/688例)。对前500例训练数据集进行随机森林模型判别,其抑郁判断的正确率为98.2%;188例测试集判断结果正确率达99.47%;将688例患者资料运用随机森林模型进行学习,总的正确率为98.84%。重要性测度结果显示,病灶位置、中医药干预手段及抑郁家族史是脑卒中后是否发生抑郁最重要指标的前3位。结论病灶位于额颞叶的脑卒中患者以及具有抑郁史的患者更容易发生卒中后抑郁。

Objective To determine the influencing factors of post-stroke depression by machine learning.Methods Stroke patients＇ medical records（ 688 cases eligible） were extracted from record system,including age,gender,pulse manifestation,complexion,tongue quality,fur,Chinese medicine intervention,body mass index（ BMI）,blood pressure,blood glucose,blood triglyceride,blood total cholesterol,smoking history,drinking history,depression family history,stroke lesion site in imaging,as well as the final depression judgment. Single rule algorithm（ 1R） was adopted to learn. The risk factors influencing post-stroke patients＇ depression in extracted information were determined. Then the cases collected were divided into the training dataset（ 500 cases） and the test dataset（ 188 cases）. Optimal discriminant results were obtained by random forest model. Results Single rule algorithm showed that the most important influencing factor of post-stroke depression was stroke lesion site. By computer speculation,stroke lesions in the frontal and temporal lobes were most prone to post-stroke depression. Basal ganglia,brain stem,cerebellum,medulla oblongata and occipital lobe lesions were less likely to cause depression. The accurate classification rate could amount to 88. 95%（ 612/688 cases）. Random forest model determination was made in the former 500 cases in the training dataset. The total correct rate of determining depression was 98. 2%. The total correct rate of determination in 188 cases of the test dataset was 99. 47%. Six hundred and eighty-eight patients＇ data were learnt by random forest model. The total correct rate was 98. 84%. The importance measure results showed that top 3 important indexes of post-stroke depression were lesion site,Chinese medicine intervention and depression family history.Conclusion Patients with lesions in the frontal temporal lobes and depression family history were most prone to post-stroke depression.