社区管理高血压患者的服药依从性预测模型构建

Developing predictive models of medication adherence for community-managed hypertensive patients

目的建立管理高血压患者服药依从性预测模型,探索预测因子对服药依从性预测能力及人群分亚组建模对预测能力的影响。方法采用问卷调查1 131名管理高血压患者服药依从性、预测因子,单因素分析筛选有意义变量(P<0.1),穷举其所有可能组合形成多个数据集,反向传播算法建模,使用十折交叉验证,计算模型符合率。归纳符合率前5的模型包含的变量且使用逐步logistic回归验证。对比将生活方式依从性、心理变量加入到模型中并对人群分组建模对模型符合率的影响。结果 17个变量中筛选出9个有意义的,产生511个数据集后建模,符合率为62.51%±0.25%。符合率前5的模型均纳入了居住地(β城市=0.457,P<0.001,社会经济)、高血压疾病严重程度(β轻度不适=-0.417,P=0.002;β中度及以上不适=0.007,P=0.974,疾病相关)、共患糖尿病(β是=0.308,P=0.019,疾病相关)、收缩压(β=-1.232,P=0.071,疾病相关)和是否享受慢病保险(β是=0.382,P=0.009,卫生系统相关)。预测因子中加入生活方式依从性后,模型符合率均增大(最大为64.46%),加入精神心理类变量后,模型符合率均增大(最大为81.96%)。血压已达标组模型符合率为79.52%,未达标组为85.01%。结论模型有一定预测能力,符合率最大为81.96%;社会经济、疾病相关和卫生系统因素有最强预测潜力;生活方式依从性和精神心理类变量对服药依从性有预测性;对样本分亚组建模不能提高模型预测能力。

Objective To develop models to predict the medication adherence for community-managed hypertensive patients,to explore the predictive ability of predictors,and to test the predictive ability of models in subgroups.Methods Data on medication adherence and predictors of 1131 respondents were obtained throught a survey.Univariate analysis was performed to select significant variables(P<0.1)to enumerate all the possible combinations of them to create datasets.Back propagation algorithm was used to develop models.10-fold cross validation was used.Accuracies of models were calculated.The variables in the models with the top 5 accuracies was summarized and tested with stepwise logistic regression.We added lifestyle adherence and psychological variables as predictors,and splited sample into subgroups to compare the change of accuracies.Results Nine significant variables were selected from 17 variables,and 511 datasets were generated for modeling,with accuracies of 62.51%±0.25%.The models with the top 5 accuracies all included residence(β_urban area=0.457,P<0.001,socioeconomic factor),the hypertension severity(β_mild discomfort=-0.417,P=0.002;β_moderate and above discomfort=0.007,P=0.974,disease related factor),whether comorbid with diabetes(β_comorbid=0.308,P=0.019,disease related factor),systolic blood pressure(β=-1.232,P=0.071,disease related factor),and whether covered by chronic disease insurance(β_covered=0.382,P=0.009,health system factor).The accuracies increased after adding lifestyle adherence(maximum:64.46%)or psychological factors(maximum:81.96%)into models.The accuracy was 79.52%for the subgroup with their blood pressure controlled and 85.01%for uncontrolled.Conclusions The models have certain predictive ability with highest accuracy of 81.96%.Socioeconomic,disease related and health system factors have the strongest predictive potential.Lifestyle adherence and psychological factors have some predictive ability.Splitting sample into subgroups cannot increase predictive power.