| Objective:The machine learning method was used to develop and verify the predictive models of medication adherence for community-managed hypertension patients.To explore the predictive ability of classification algorithm,predictors,lifestyle adherence,and developing models in subsamples on predicting medication adherence.Methods:Using typical sampling,we conduct a cross-sectional survey in the Xinjiang production and construction corps in July 2017.3 communities in the city and 2 rural areas were selected and a face-to-face survey was used to collect data.Adherence,predictors,and life-style adhernce were measured using a self-conducted questionnaire.Univariant analysis was used to select the predictors that have a statistic relationship with medication adherence.Then all possible combinations of the selected variables were generated.Support vector machine,random forest and back propagation algorithm were used for developing predictive models.The performances of the models were evaluated by ten-fold cross validation and were assessed with accuracies.The five models with the highest accuracies in each kind of algorithm were considered as the 5 best models.The predictors included in these 15 models were analyzed to find out the categories of variables that may have a potential to predict adherence.Adherence to recommend lifestyle was add to the best15 models to compare the changes of the accuracies.The support vector machine model,random forest model and back-propagation algorithm model with the highest accuracies were selected as the three best models.The patients were divided into two subgroups according to whether their blood pressure under control,and the three best models were redeveloped in subsamples to examine whether the performances varied in the different subgroups.Results:1.A total of 1131 eligible hypertensive patients were included in this study.The age of the patients was 73.13 ± 7.40 years old,and females accounted for 63.13%.Systolic blood pressure was 136.95 ± 15.78 mmhg and diastolic blood pressure was 77.66 ± 10.10 mmhg.There were 664 patients who had their blood pressure under control and 426 patients with a poor medication adherence.A total of 9 variables had a statistically significance in the distributions among two groups: self-rated health,income,residence,disease severity,co-diabetes,duration,and outpatient medical reimbursement.In addition,systolic and diastolic blood pressure were included based on the experience.A total of 9 predictors was selected.2.511 possible combinations were generated base on 9 selected predictors.Along with support vector machine,random forest,back propagation algorithm,1533 models were generated.The accuracies ranged from 56.06 percent to 63.48 percent(61.76% ± 1.04%).The average accuracy of SVM model was 62.20 ± 0.003%,the average accuracy of random forest model was 60.88 ± 1.33%,and the average accuracy of back-propagation algorithm was 62.21 ± 0.004%.3.The five models with the highest accuracies for each kind of models were selected.Socio-economic,health system related,and condition-related factors were the most frequently included variables in the 15 highest-accuracy models.4.When adherence to recommended lifestyle was added as a predictor in the above 15 models,the models showed the highest accuracies.The accuracy was 64.10% for the best support vector machine model,63.40% for the random forest model,and 64.46% for the back-propagation algorithm model.5.The patients were divided into the uncontrolled group(N = 467)and the controlled group(N = 664).Compared with the back-propagation algorithm model developed base on overall samples,the accuracy in the controlled group increased from 63.46% to 64.91%.The accuracies of other models developed base on the subgroup samples decreased.The accuracy of the best support vector machine models decreased to 60.17% in the uncontrol group and 63.25% in the controlled group.The accuracy of the best random forest models decreased to 62.53% in the uncontrolled group and 59.64% in the controlled group.The accuracy of the best back-propagation algorithm model decreased to 63.60% in the uncontrol group.Conclusion:The models for predicting medication adherence of community-managed hypertensive patients has some predictive ability.Socio-economic factors,health system related factors,and condition-related factors have the best predictive potential.Adding lifestyle adherence as a predictor to models can increase the predictive ability to some extent.In this study,dividing patients into subgroups according to whether their blood pressure under control did not increase the predictive ability of the models. |
PDF Full Text Request