This cross-sectional descriptive survey was undertaken at outpatient clinics at a teaching hospital in Cheonan, Korea. We used sample data from January to May 2011 of 293 patients over 65 years of age with chronic disease. We included elderly patients who had been taking a medication for longer than 6 months and had asthma, hypertension, diabetes, chronic obstructive pulmonary disease, liver cirrhosis, stroke, and cardiovascular diseases with normal cognitive function. The study was approved by the ethics committee of the hospital prior to the start of data collection. Written consent was obtained. The questionnaire was verbally administered to consenting respondents who were unable to self-complete the survey.

Sixteen variables were used: age, gender, job, monthly income, spouse, educational level, activities of daily living (ADL), perceived health status, duration after diagnosis, number of medication types, daily pill counts, side effects of medication, self-efficacy, depression, health literacy, and medication knowledge. The variable sets used in this study are shown in Table 1.

The questionnaire was designed to yield information about demographic characteristics such as age, gender, educational level, spouse, and monthly income. Depression was assessed by means of the short-form of the Geriatric Depression Scale (GDS), which is a validated 15-item, self-report depressive symptom scale designed to detect the presence of current depression in older adults [11] (Cronbach's α = 0.91). To assess patients' medication knowledge, we used five self-report questions using a 5-point Likert scale [12] (Cronbach's α = 0.79). Health literacy indicates an individual's ability to obtain and use health information to make appropriate decisions for health and medical care. For health literacy, patients were asked three previously described health literacy screening questions, each with five possible response options [13] (Cronbach's α = 0.81). Perceived health status refers to self-reported status, ADL and daily self-care capabilities at home and in outdoor environments, respectively. In this study, self-efficacy indicated a patient's belief in his/her ability to succeed in adhering to the prescription medication. We used a previously described medication self-efficacy system [14] using 13 items with good internal consistency reliability (Cronbach's α = 0.89).

Lastly, medication adherence was determined using a modified version of the four items, self-reported Morisky medication adherence scale [15]. Each item was in a yes/no format with each item is in a yes/no format with a maximum possible score of four equating very poor adherence and 0 or 1 typically considered as good adherence. The Morisky scale has been used across many chronic diseases, as a self-reported measure of adherence to medications and has demonstrated good reliability and predictive validity [16]. Scores ≤2 were considered indicative of non-adherence to medications.

Sixteen variables were used in the model building and analysis. The variables were selected because they either had been shown to have an impact on medication adherence in previous research [17] or were of potential clinical importance as indicated by a panel of experts. The variables were gender, age, job, educational level, side effects of medication, depression, health literacy, monthly income, spouse, and duration after diagnosis, medication knowledge, number of medication types, daily pill counts, perceived health status, ADL, and self-efficacy. Patients were asked about medication adherence using a self-reported questionnaire. The dataset was divided in a group of 120 adherent patients and 173 non-adherent patients.

The LR model building processes were carried out in SPSS ver. 20.0 (IBM, Armonk, NY, USA). A p-value ≤ 0.05 was considered to be significant for inclusion into the model. We checked multicollinearity to verify the adequacy of the regression model. The identified variation inflation factor (VIF) value of less 10 confirms the absence of collinearity (in the data). Moreover, the Durbin-Watson test for residual analysis showed a value of less than 2 (1.85), effectively demonstrating that there was no correlation between the error terms of the model, which subsequently satisfied the assumption of normal distribution of residuals. Therefore, it entailed that the data under study was suitable for regression analysis.

The SVM-based model building processes were carried out with MATLAB ver. 7.12 (MathWorks, Natick, MA, USA). We used SVM with radial basis function (RBF) as kernels.

Comparison of LR and SVM discrimination for both models was performed. Five widely used statistics were adopted to evaluate the performance of a model: sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy. To test the ability of each model to distinguish patients, the area under the receiver operating characteristic curve was calculated.

We used sample data of 293 patients with chronic disease (120 good adherence and 173 poor adherence results). The mean age of the patients was 73.8 years. Table 2 shows the socio-demographic and clinical characteristics of the patients.

Taking into account the 16 variables, the results of applying LR accuracy was 71.1%. Duration after diagnosis and self-efficacy are selected as significant variables in the LR model. The medication adherence rate of patients with duration after diagnosis of more than 10 years was 46% lower than those with duration after diagnosis of less than 5 years. For every unit higher in the self-efficacy score, the medication adherence rate rose by 27%. A complete list of study variables in each variable set along with p-values are listed in Table 3.

To examine the characteristics of the patient samples with good and poor prognoses before performing the SVM experiments, we applied Principal Component Analysis (PCA). PCA transforms original features in a multivariate data set into salient features that are not correlated with each other [18]. Therefore, the original features, representing the patient sample, can be reduced to a smaller number of new features, referred to as the principal components (PC). The largest variance for the data set is set as the first axis (the first PC) in the coordinate system. Likewise, the second greatest variance is set as the second axis (the second PC), and so on. Applying PCA to the 16 features of the 293 patient samples, we could project the patients onto a three-dimensional space composed of PCs 1, 2, and 3. Figure 1 illustrates the distribution of the patient samples on this coordinate. From this figure, we expected to develop a SVM classifier that distinguishes the patients with good and poor samples.

To identify the variables that had the highest classification accuracy in medication adherence for chronic disease, we developed SVM with radial basis function (parameter C = 1, γ = 1/number of features) that systematically searched through the space of subsets of variables, and evaluated the goodness of each variable subset according to the prediction accuracy. The variable subset showing the highest accuracy was identified as the predictor set. Parameter C is the weights between empirical error and generalization error. Parameter γ controls the shape of the separating hyperplane.

Although an exhaustive enumeration search of variables can find an optimal solution, it requires extremely high computational cost to train and test SVM with each subset of variables. Thus, we employed sequential forward selection (SFS) search to deal with this difficulty [19]. The SFS is a heuristic greedy search that starts from an empty set of variables, sequentially selecting a variable.

We listed the top nine ranked variables selected by SVM and their prediction accuracies using a combination of the top ranked variables together in Table 4 to examine the above results in detail. The accuracy using a single variable selected was 72.4%; self-efficacy was selected, as in the LR model. The present accuracy of the SVM reached 83.3% with two variables, self-efficacy and age. The highest accuracy, 97.3%, was achieved with nine predictors: self-efficacy, age, depression, health literacy, medication knowledge, number of medication type, daily pill counts, duration after diagnosis, and education level.

Figure 2 shows the prediction accuracy when all 16 variables were used for the prediction of medical adherence in the order of variables selected by SVM. The performance was very markedly decreased when more than 10 features were selected. Unlike our intuition that having more variables should give higher predictive performance, this example demonstrates that using a small number of variables can achieve higher prediction accuracy.

Table 4 compares the experimental results of LR and SVM using five evaluation measures. SVM showed better performance than LR in overall scoring categories, allowing identification of predictor candidates to determine the most probable medication adherence of a patient.

LR showed 71.1% accuracy when all 16 variables were used for the prediction of medical adherence (Tables 3 and 5). Compared to the result of LR, the result of SVM showed significantly higher accuracy, 97.3%, with only nine variables on the same patient samples (Table 5). This result indicates that SVM can achieve greater accuracy with a smaller number of variables than the number of variables used in LR. It is interesting to note that the most significant variable (self-efficacy) selected by the SVM agrees with that selected by LR. When even a single variable (self-efficacy) was used by SVM, 72.4% accuracy could be achieved, which is higher than that achieved using all variables by LR.

The results of the comparison of the discriminatory power of LR and SVM models are summarized in Table 5.

The AUC indicates how well a prediction model discriminates between healthy patients and patients with disease. The following guidelines have been proposed for interpretation of this area: 0.5-0.7, rather low accuracy; 0.7-0.9, moderate accuracies useful for some purposes; and >0.9, rather high accuracy [20]. Therefore, the classification accuracy of these models was good.

Our results indicate that the SVM model has better diagnostic capability than LR model. The AUC has achieved a good diagnostic power. Figure 3 shows the performance of the two models built.