Identification of predictive factors of the degree of adherence to the Mediterranean diet through machine-learning techniques

Population and data description

This is an observational prevalence study, conducted in Northwestern Spain (municipality of Cambre, A Coruña, Spain), which included a randomly selected population aged 40 years and over. The sampling population consisted of individuals residing in Cambre, identified through the National Health System card census. In Spain, the National Health System has universal coverage, and almost all Spanish citizens are beneficiaries of public healthcare services.

The sample size was calculated taking into account the total population of the municipality ($n=12,446$). After stratification by age and gender, ($n=503$) persons were selected to participate in the study. Sample size was estimated using the single proportion formula, with 95% confidence Interval. A sample size of ($n=503$) subjects was estimated based on an adherence to MD rate of 50%. Precision was set at 4.3% and percentage of losses at 10%. Population data is shown in Table 1.

A personal interview was arranged with each individual. After obtaining their permission and written consent, a trained nurse proceeded to the measurement of anthropometric variables and to the collection of the necessary data to cover the questionnaires. The patients who could not go to the health centre due to personal or displacement reasons and those who suffer from a cognitive impairment, making it impossible for them to perform the study, were excluded. The study received written approval from the Regional Ethics Committee for Clinical Research (code 2012/390 CEIC Galicia).

The information described below was collected from each selected subject: socio-demographic variables: age, gender, level of education, marital status and relationships of coexistence; prevalence of arterial hypertension and smoking: the systolic and diastolic blood pressure of each patient was recorded at the beginning and at the end of the visit, obtaining the prevalence of arterial hypertension; the smoking habit was recorded according to self-reported information. Anthropometric variables: the anthropometric parameters allow us to know the state of the protein and caloric reserves, besides providing guidance to the health professional about the consequences of the imbalances in these reserves.

All measurements were made during the same session, to avoid variations in the environmental or biological conditions. For the measurement of weight and size, the person was barefoot and with light clothing; an MB-201T plus Asimed scale-rod was used with an accuracy of 100 grams (weight) and 1 mm (size). BMI was obtained by means of the $\mathrm{B}\mathrm{M}\mathrm{I}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}={\displaystyle \frac{\mathrm{w}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}(\mathrm{k}\mathrm{g})}{\mathrm{h}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{t}({\mathrm{m}}^2)}}$, and grouped according to the WHO classification of $\mathrm{B}\mathrm{M}\mathrm{I}<{\displaystyle \frac{18.5\mathrm{k}\mathrm{g}}{{\mathrm{m}}^2}}$: low weight; 18.5–24.99 ${\displaystyle \frac{\mathrm{k}\mathrm{g}}{{\mathrm{m}}^2}}$: normal weight; 25–29.99 ${\displaystyle \frac{\mathrm{k}\mathrm{g}}{{\mathrm{m}}^2}}$: overweight, and $\ge 30{\displaystyle \frac{\mathrm{k}\mathrm{g}}{{\mathrm{m}}^2}}$: obesity.

The waist and hip circumference was measured with an inelastic tape measure with the patient standing upright, the abdomen relaxed, the upper limbs hanging at the sides, and with the feet and knees joined together. The waist circumference was measured by taking the mid-point between the lower costal margins and the iliac crests, as it is considered a risk factor for CVD when it is wider than 80 cm in women and wider than 94 cm in men, and a very high risk if it exceeds 88 cms and 102 cms, respectively (Alberti et al., 2009).

The hip circumference was measured as the maximum circumference around the buttocks. Based on these two values, the waist-hip ratio was calculated using the cut-off points proposed by the WHO, where normal levels of 0.8 are found in women and one in men, higher values indicating abdominal visceral obesity, which is associated with increased cardiovascular risk (Jover, 1997).

The calf circumference was measured in the widest section of the ankle-knee distance (cuff area) showing a good correlation with fat-free mass and muscle strength (Rolland et al., 2003; Barbosa Murillo et al., 2007; Bonnefoy et al., 2002). The measurement was carried out with an inextensible tape measure in cm.

Subscapular skin fold, this fold measures truncal obesity. The measurement is made one centimetre below the lower angle of the scapula, following the natural furrow of the skin. The scapula protrudes when the arm is carefully placed behind the back and the lower angle can be located this way. The measurement of the fold will be diagonally over an angle of 45° to ensure the correct thickness measurement. The plicometer forceps should be applied 1 cm in the inferolateral position to the thumb and finger that lifts the fold.

To assess the amount of subcutaneous adipose tissue, the skin folds were measured in millimetres in the tricipital, bicipital, subscapular and suprailiac areas. A digital calliper Trim metre was used, including a double layer of skin and underlying adipose tissue, always avoiding the muscle. The tricipital skinfold was measured longitudinally, at the back of the non-dominant upper limb, at the midpoint between acromion and olecranon, with the limb relaxed, parallel to the axis of the arm; the bicipital fold was measured at the same point as the tricipital, but on the under arm.

The circumference of the arm was measured with an inextensible anthropometric tape measure in cm. The measurement was taken at the midpoint of the non-dominant arm, in the same place where the tricipital skinfold was measured and without compression with the anthropometric tape.

Once the data of the different measurements were obtained, the mid-arm muscle circumference was found, with which the skeletal muscle mass of the patients (protein compartment) was known and expressed in cm. The arm muscle area indicated that the muscle compartment was based on the brachial circumference and tricipital skinfold measurements. The fat area of the arm indicated that the patient’s fatty compartment used the total brachial area and the muscular area of the arm. The Adipose Muscular Index, which evaluates the nutritional status from the adipose and muscular areas of the arm, was also calculated, being essentially applied in the assessment of obesity.

For the determination of body fat percentage by electrical bio-impedance, a Beurer and BG55 model bio-impedance metre was used, with a maximum capacity of 150 kg and a precision of 0.1% for body fat, body water and muscle percentage, and 100 g for body weight, according to the information provided by the manufacturer.

These methods are based on physical principles, such as the different ability of conduction or resistance that the tissues show to the passage of an electric current, with greater conductivity of the lean tissues than the fatty ones (Norman et al., 2007). Thus, by means of bio-impedance, the following values were obtained: weight, fat mass, liquid mass, muscle mass, bone density, basal metabolic rate (BMR) and activity metabolism. Data on socio-demographic variables, such as age, gender (male/female), cohabitation (with whom the live), prevalence of current smokers, ex smokers (patient stopped smoking more than 12 months before entering the study) and non-smokers were estimated. Additionally, blood pressure was recorded.

Adherence to the MD

Consumption of a characteristic food pattern of MD is associated with numerous health benefits. These benefits are attributed to bioactive compounds that exert synergistic effects and decrease the risk for development of chronic diseases.

In order to assess the quality of dietary habits (adherence to a Mediterranean dietary pattern), the MD adherence test was used (Estruch et al., 2018). It is a questionnaire consisting of fourteen quick questions that allow us to understand whether participants’ usual diet can be considered as following the parameters of the MD. Each question answered affirmatively adds a point. It is considered that a person correctly follows the MD when their score is equal to or greater than nine points.

The assessment of nutritional status was determined using the Mini Nutritional Assessment (MNA) questionnaire (Estruch et al., 2018). It is a validated method which, through eighteen short questions, evaluates anthropometric measures, dietary habits, lifestyle, pharmacological treatments and mobility, and performs a subjective evaluation of health and nutritional status. The total value of the MNA scale is thirty points, a score $<17$ being considered malnutrition, there is a risk of malnutrition between 17 and 23, 5, and well nourished subjects obtain scores of 24 points and higher.

A measure of subjective weight is included by asking: ‘I consider that my weight is: (A) higher than normal, (B) normal, (C) lower than normal’, following the model proposed in (Espina et al., 2001). Based on the answer, the population is classified into three groups: ‘fairly subjective weight’ those who believe to be at an ideal weight, ‘more subjective kilograms’ for those who believe that they are overweight and ‘less subjective kilograms’ for those who think they weigh less than they should.

Two of the eleven sub-scales of Garner’s Eating Disorder Inventory (EDI-2) (Garner, 1998) were used to study body image: Body Dissatisfaction (EDI-IC) and Obsession for Thinness (EDI-OD), as they evaluate aspects directly related to perceptual alterations. The body dissatisfaction sub-scale (EDI-CI) measures the dissatisfaction of the subject with the general shape of their body or with those parts of the body that most concern those with eating disorders (stomach, hips, thighs, buttocks). The thinness obsession sub-scale (EDI-O) measures concern about weight, diets and fear of getting fat.

This questionnaire was validated in Spain by (Corral et al., 1998). The fourteen items of these two sub-scales were mixed in the questionnaire to avoid the subjects guessing the construct being evaluated. All items were answered and corrected according to the form proposed in the questionnaire manual. The MD Adherence test was used to determine the degree of adherence to the MD, being a short specific questionnaire of fourteen items validated for the Spanish population and used by the MD Prevention Group (PREDIMED) (Martínez-González et al., 2015).

Machine learning and statistical analysis

The authors tested different ML techniques for solving this problem, using cross-validation techniques to avoid over-training, while ensuring that the generalised capability of the model is the best possible, as well as different runs of the experiments to check the behaviour of the techniques. Thus, all experiments were repeated ten times to check the stability of the results and the observed deviation between the experiments was small, as shown in the results section. In particular, a tenfold cross validation was used to divide the dataset in such a way that nine random partitions were used to train and one to validate the results, each time taking a different subset for validation.

In order to compare the performance of the ML techniques, the Area Under the Receiver Operating Characteristic Curve (AUROC) was used. This is a combined measurement which, besides being independent of the threshold used, includes both Type I and type II errors, ensuring that it is not conditioned by differences in the total number of cases of each class (Fawcett, 2006).

An experimental design was employed (Fernandez-Lozano et al., 2016a), allowing us to divide the data using a cross-validation technique which ensured that the performance results obtained, as mentioned above, were not skewed. That is they were adjusted to the data, and researchers are able to identify which of the hyper-parameters are most suitable to find the best model with each ML techniques, according to its particular hyper-parameter configuration. To this end, the programming environment R (R Core Team, 2016) and the package mlr (Bischl et al., 2016) were used, which also allowed us to perform the considered experimental design. In addition, another of the objectives pursued by this study was to find as few variables as possible that would yield a performance value as high as possible, preferably at least equal to that obtained using all available variables. This is basically a feature selection approach where the main aims are the following: avoid overfitting and improve model performance, to provide faster and more cost-effective models, and moreover to gain a deeper insight into the underlying processes that generated the data as mentioned in (Saeys, Inza & Larrañaga, 2007). There are three approaches in ML to perform this process and the use of a filter approximation was chosen, for its velocity and independence of the classifier (Saeys, Inza & Larrañaga, 2007). In general, performing this feature selection process helps to reduce inherently the present noise in such datasets.

The final step of our experimental analysis was a null hypothesis test for choosing the best model in order to ensure whether the performance of a particular ML technique is statistically better than the others or not. In our case, as there were more than two repeated measures, an ANOVA or a Friedman test should be considered. In particular, three different conditions should be checked: normality, independence and homoscedasticity. If our results fulfil the three conditions, a parametric test is applicable, and the ANOVA one should be considered, otherwise the non-parametric version, the Friedman test. Finally, a post hoc procedure had to be used in order to correct the p-values for multiple testing.

Machine learning techniques for classification problems

A large number of experiments were carried out in an attempt to identify the best ML model able to solve the problem and to ensure that the results are reproducible, real and obtained under equal conditions. In addition, the search space was explored for the best possible parameters for each technique in the same way, so that all techniques could have the same possibilities of exploration across the same subsets of data and avoid the over fit that could occur. In particular, the following well-known state-of-the-art techniques were implemented: Random Forest (RF) (Breiman, 2001), Support Vector Machines (SVM) (Cortes & Vapnik, 1995; Vapnik, 1995), Elastic Net (ENET) (Tibshirani, 1994; Zou & Hastie, 2005) and weighted k-Nearest Neighbours (KNNs) (Hechenbichler & Schliep, 2004).

Random Forest (Breiman, 2001) is a state-of-the-art ML technique that was used in multiple domains with good results. One of its main strengths is that the results obtained are very easy to understand, it is based on very simple concepts and in general, although it is applied with little experience in the parametrisation of hyper-parameters, good results are obtained. This technique combines multiple decision trees, each of them tuned over a subset of bootstrapped data. In this way, RF combines each of the individual predictions of the decision trees into a global prediction that, in general, is more successful than any of the simple ones. Of all the possible variables in the dataset, a number were randomly selected (with replacement) and a number of trees were constructed based on the set of examples used for the training phase and obtained from the previously selected subset. When there are classification problems, it is recommended to use a square root number of the total number of variables existing in the dataset. To explore the solution space in the best way possible, in our experiments we used a parameter domain that was adjusted by a grid search and that, for a number of trees (1,000), we explored randomly selected values of variables (2–6). In addition, values that varied (1–4) according to the size of the terminal nodes of the tree were explored.

Support Vector Machines (Cortes & Vapnik, 1995; Vapnik, 1995) is also one of the ML techniques that have been commonly used in different domains in recent years and have obtained good results. In fact, along with RF, it is one of the algorithms considered state-of-the-art, easy to understand, and the results obtained are verifiable. In problems that occur during a study, the main objective of SVM is to find the hyperplane that best separates the examples between high and low degree of adherence to the MD and at the same time to maximise the distance of separation between both examples and the hyperplane. That is it attempts to find the separation hyperplane that generalises in the best possible way (Burges, 1998). To achieve this goal, SVM introduces a particular mathematical concept known as kernel: it is a mathematical function that allows the conversion of the input space into a higher dimension, which is used to transform a non-separable linear problem into one that is separable. There are different kernel functions, which in general could be interpreted as a measure of similarity between two objects (60), and one of the most used is Gaussian Radial Basis (RBF), because basically any surface can be obtained with this function (61). In this case, the domain of the parameters used to search for the best model consists of a grid search of two different parameters. The first one (parameter C) is directly related to the model and is used as a balance between the classification errors and the simplicity of the decision surface, while the second (gamma parameter) is the free parameter of the Gaussian function and in particular, SVM is very sensitive to changes in this parameter. For both parameters, and according to the usual practice, values were evaluated in potencies of two between −12 and 12. To better understand this technique, the following reading materials are recommended (Burges, 1998; Vert, Tsuda & Schölkopf, 2004; Cristianini & Shawe-Taylor, 2000).

Elastic Net (Tibshirani, 1994; Zou & Hastie, 2005) is based on lasso (penalised least squares method) and was specifically developed to solve some of the limitations encountered for this technique (56). On the one hand, a grid search was performed on two different parameters, the alpha penalty parameter was searched (it has values in the range of 0–1) and in particular the following 0, 0.15, 0.25, 0.35, 0.5, 0.65, 0.75, 0.85 and 1. On the other hand, the best value of the lambda parameter was used, as recommended by the authors of the technique, from values less than or equal to one to negative powers of ten, in particular the following values were used: 0.0001, 0.001, 0.01, 0.1, 1.

Finally, a simple KNN (Hechenbichler & Schliep, 2004) assigned, through a decision rule, an unclassified example belonging to a class by frequency of occurrence to its k-most similar neighbouring examples. Then, in accordance to the distance of Minkowski for each of the examples and following the maximum accumulated kernel densities the weighted KNN are identified (Hechenbichler & Schliep, 2004; Samworth, 2012). In particular, neighbouring values of less than or equal to nine were used. Therefore, this particular and improved implementation of a KNN used kernel functions to measure the degree of similarity of the examples, as previously mentioned in the case of the SVMs.