Multilayer perceptron modeling of health literacy and medication adherence determinants in Arab patients with cardiometabolic disease: the role of quality of life, eating behaviors, and physical activity

Abstract

Objective

Cardiometabolic diseases (CMDs) are a leading cause of global mortality, with an increasing prevalence in the Arab world, particularly among older patients. This pioneer study using the MLP modeling technique will investigate the determinants of health literacy and medical adherence among Arab CMD patients, focusing specifically on quality of life, eating behaviors, and physical activity. Here we conducted study on Arab patients with cardiometabolic diseases using validated instruments to assess medication adherence, health literacy, quality of life, eating behavior, and physical activity and observed the relationships between them using machine learning modelling technique.

Results

The study demonstrated that 82.4% of participants reported low medication adherence, and 39.8% had adequate health literacy. Physical activity had a weak positive correlation with medication adherence while participants with higher adherence were somewhat more health-literate but reported lower enjoyment of food. The study observed using the MLP model that the medical adherence and the health literacy models demonstrated strong classification performance for predictors like quality of life, hunger, emotional overeating, and satiety responsiveness. The study highlighted that the ML model identified quality of life, eating behavior, and physical activity as key predictors for health literacy and medication adherence among Arab cardiometabolic patients.

Introduction

The increase in obesity, coronary diseases, metabolic syndrome, and diabetes over recent decades has been associated with extensive lifestyle changes in the Middle East region [1]. These factors have led to decreased physical activity, while the excessive consumption of processed, calorie-dense foods with little nourishing value has also become more common [2, 3]. These risk factors result in the significant rise of obesity and cardiovascular conditions among patients in the Middle East [4, 5]. In 2013, the WHO indicated that at least 80% of cardiovascular and metabolic disorders could be prevented through secondary prevention strategies [6]. The alternative to obesity is a healthy lifestyle, which includes regular physical activity and dietary modifications [7]. A study in 2024 noted that about 20% of patients were readmitted within three months, and nearly 40% within a year, mainly due to poor adherence to medication, diet, and physical activity [9]. Adopting healthy behaviors improves quality of life, reduces recurrent cardiovascular events, and lowers readmissions [10]. Research has linked non-adherence to worse cardiovascular outcomes, poorer health, and higher mortality [11–16].

Health literacy (HL), the ability to find, understand, and use health information for decisions, is crucial for adherence [17–19], following oral health instructions, informed decision-making, and timely advice-seeking [20, 21]. Although extensive research has been conducted globally on medication adherence among patients [23, 24], including systematic reviews that identify various factors related to adherence, studies on lifestyle adherence among CMD patients in the Arab region are limited [8]. Moreover, conventional modeling techniques such as regression and logistic models are widely used to identify factors linked to health outcomes. These methods often assume linear relationships, which may not capture the complexity and non-linearity of real-world data. In contrast, this study employs a machine learning approach, specifically a multilayer perceptron, to model complex, non-linear relationships between predictors and outcomes. Using this model to identify low adherence, quality of life, and related factors among older Arab patients, this approach has the potential to enhance the accuracy of modeling patient outcomes.

Therefore, this study is focused on evaluating medication adherence and HL levels among cardiometabolic patients, classifying quality of life, sociodemographic, and eating behavior factors that can affect adherence and HL, and finally determining the important predictive factors.

Materials and methods

Study design

A cross-sectional design was used among Arab patients suffering from CMDs. Three Different hospitals, both government and private, were selected for the sample calculation. Data were collected at a single point in time to assess the current status of medical adherence and health literacy among participants. The convenience sampling technique was adopted to recruit patients. The choice of sampling methodology was based on specific conditions, as the collection of such samples from elderly patients from tertiary hospital settings was challenging. G-Power software version 3.0.10 was used to estimate using α = 0.05, 1 - β = 0.85, resulting in a sample of 269 participants [22].

Ethical considerations

This study has been approved by the Institutional Review Board (IRB), KACST, KSA, with the ethical no (HA-01-R-104).

Inclusion criteria

The inclusion criteria of participants were required to be 18 years or older, diagnosed with CMDs (e.g., cardiovascular disease, type 2 diabetes mellitus), and self-identify as Arab. Moreover, they are required to have the ability to understand sufficient English or Arabic to complete the study instruments, ensuring accurate and reliable data collection. The data collection was also conducted face-to-face, and assistance was provided to our older age group participants to ensure accurate completion of the questionnaires.

Variables in the study

The study was conducted using a self-report questionnaire, which included demographic information such as age, gender, marital status, BMI, work status, and level of education. The study used the MARS5 for medical adherence, HL-EU-Q16 for health literacy, GLTEQ for physical activity, AEBQ for eating behavior, and QoL for assessing quality of life. They are discussed in detail below.

Instruments of the study

MARS-5

MA was assessed using the 5-item MARS, a validated self-report measure designed to evaluate both unintentional (e.g., forgetting) and intentional (e.g., altering or skipping doses) non-adherence. For this study, adherence was categorized into three levels based on total MARS-5 scores: scores of 15 or below were classified as low adherence, scores between 16 and 20 were considered moderate adherence, and scores of 21 or higher were categorized as high adherence. The version of MARS-5 was used with permission [44].

HL-EU-Q16

The HLS-EU-Q16 is a 16-item self-report questionnaire designed to assess the challenges individuals face in accessing, understanding, evaluating, and applying health information for decision-making in the areas of health care, disease prevention, and health promotion [34].

GLTEQ

Physical exercise was assessed as an indicator of the adoption of healthy lifestyle behavior using the Godin Leisure Time Activity scale [25]. It is a self-administered 7-day recall questionnaire and has 3 levels of exercise: strenuous, moderate, and mild.

AEBQ

The AEBQ assesses eight appetitive traits, categorized into food-approach behaviors (e.g., responsiveness to food cues, emotional overeating) and food-avoidance behaviors (e.g., satiety sensitivity, food selectivity), as validated by [26].

QoL

The QoL scale used in this study was categorized into three levels, Low, Moderate, and High, based on standardized scoring procedures adapted from instruments such as the WHOQOL-BREF. Scores ranging from 0 to 49 were classified as indicating low quality of life. Scores between 50 and 74 were considered moderate. Scores from 75 to 100 were categorized as high. This study assessed quality of life using the WHOQOL-BREF, a validated, freely available instrument developed by the World Health Organization [43].

MLP modeling

The MLP is a powerful type of machine learning approach designed to model complex, non-linear relationships within health-related data. This model will examine the key factors contributing to medical adherence and health literacy among patients. MARS and HL will be treated as dependent variables, and the MLP will evaluate the significant predictors influencing their roles in the prediction model. Specifically, the MLP model used normalized importance as a measure of feature importance based on the analysis of connection weights within the neural network. This method will assess the relative contribution of each input variable to the network’s output by analyzing the strength of connections from inputs to the hidden layer and then to the output layer.

Statistical analysis

The reliability of the questionnaire was evaluated by Cronbach’s alpha, a statistical measure that assesses the internal consistency of the instrument. Subsequently, descriptive analysis involving frequencies, percentages, and mean ± SD was calculated for the study variables. Pearson’s correlation examined the direction and strength of the relationships presented through a heatmap. The study employed the χ2-test of independence to test proportions among different categories. Data were analyzed using SPSS, version 27, and a machine learning model (MLP) was performed using R (version 4.5.1) within R-Studio Desktop (version 2025). The results were declared statistically significant if p < 0.05. Additionally, the findings of p-value and 95% CIs were supported by the effect size,

Which was computed by Cramer’s V, commonly used for chi-square tests with more than 2 2 contingency tables. Effect sizes computed were categorized as 0.1–0.2 for weak, 0.2 to 0.4, moderate while 0.6 to 0.8 for strong association. The formula for Cramer’s is:

1

Results

A total of 269 respondents with CMD participated in the study, with a mean age of 50.22 ± 14.49 years, while the BMI was 30.46 ± 9.94 kg/m². The study included 139 males (52.5%) and 126 females (47. 5%), and most patients were married (64.2%). The majority of patients were associated with King Fahd Military Hospital (KFMC), comprising 148 (55. 0%) of the sample, followed by KAAUH with 85 (31.6%). Among the participants, a notable portion, 210 (78.4%), were non-university graduates. (Table 1).

Table 1.

Baseline characteristics of cardiovascular patients

Characteristics	Categories	n	%
Age (years) 50.22 ± 14.49
BMI 30.46 ± 9.94
Hospital	KFMC	148	55
	KAAUH	85	31.6
	ALEHSAA	36	13.4
Gender	Male	139	52.5
	Female	126	47.5
Marital Status	Single	42	15.7
	Married	172	64.2
	Divorced	35	13.1
	Widowed	19	7.1
Education	University graduate	58	21.6
	Non-university graduate	210	78.4

The study indicated that 136 (50.7%) CMD patients had been diagnosed for less than 5 years, 64 (23.9%) for 5 to 10 years, and 162 (60.4%) were taking more than one type of medication. Regarding the medical history of the patients, 165 (64.5%) had hypertension, diabetes, or multiple health concerns.

Considering the physical activity, the study noted that the majority were classified as active (162 participants, 71.4%). Although medication adherence was notably low, with 215 (82.4%) and only 3(1.1%) showing high adherence, with significant differences (χ² = 291.80, p < 0.001, v = 0.74).

The majority of patients reported high enjoyment of food (183, 81.0%) while moderate food fussiness tendencies were most common (186, 71.5%), with fewer showing high levels (χ² = 140.52, p < 0.001). Nearly half of the participants of the study (n = 109, 41.5%) reported high QoL (χ² = 94.2, p < 0.001, V = 0.43). Emotional overeating and undereating were observed mainly moderate (104, 39.5%, χ² = 7.00, p = 0.030), 157, 59.5%, and high in 76 (28.8%) (χ² = 50.34, p < 0.001) among the patients, respectively. Food responsiveness, hunger, satiety responsiveness, and slowness in eating were observed mostly in moderate categories [Table 2].

Table 2.

CMD patients distribution of GLTE, MARS, HL, and AEBQ (chi-square test of proportions)

Scales	Category	N (%)	χ²-statistic (p-value)	Effect Size (Cramer’s V)
GLTE	Insufficiently Active/Sedentary	41 (18.1)	166.5 (p < 0.001***)	0.56
	Moderately Active	24 (10.6)
	Active	162 (71.4)
MARS	Low Adherence	215 (82.4)	291.8 (p < 0.001***)	0.74
	Medium Adherence	43 (16.5)
	High Adherence	3 (1.1)
HL	Inadequate HL	23 (8.6)	82.7 (p < 0.001***)	0.39
	Low HL	87 (32.3)
	Sufficient HL	107 (39.8)
	Excellent HL	52 (19.3)
AEBQ Enjoyment of Food	Low	3 (1.3)	270.9 (p < 0.001***)	0.72
	Moderate	40 (17.7)
	High	183 (81.0)
AEBQFood Fussiness	Low	16 (6.2)	140.5 (p < 0.001***)	0.51
	Moderate	186 (71.5)
	High	58 (22.3)
AEBQ Emotional Overeating	Low	99 (37.6)	7.0 (0.030**)	0.11
	Moderate	104 (39.5)
	High	60 (22.8)
AEBQ Emotional Undereating	Low	31 (11.7)	50.3 (p < 0.001***)	0.41
	Moderate	157 (59.5)
	High	76 (28.8)
AEBQ Food Responsiveness	Low	8 (3.8)	173.7 (p < 0.001***)	0.58
	Moderate	147 (70.7)
	High	53 (25.5)
AEBQ Hunger	Low	33 (15.6)	35.4 (p < 0.001***)	0.37
	Moderate	116 (54.7)
	High	63 (29.7)
AEBQ Satiety Responsiveness	Low	23 (8.8)	65.0 (p < 0.001***)	0.50
	Moderate	154 (58.8)
	High	85 (32.4)
AEBQ Slowness in Eating	Low	59 (22.5)	24.3 (p < 0.001***)	0.30
	Moderate	138 (52.7)
	High	65 (24.8)
QoL Total Score	Low QoL	47 (18.0)	94.2 (p < 0.001***)	0.43
	Moderate QoL	106 (40.5)
	High QoL	109 (41.5)

Χ²-statistic test of equality of proportions ***p < 0.001, **p < 0.05

Correlation study of HL, MARS, AEBQ, QoL, and baseline characteristics

The correlation analysis showed a positive association between age and exercise frequency (r = 0.275, p < 0.001), while age was negatively correlated with enjoyment of food (r = -0.231, p < 0.001), food responsiveness (r = -0.170, p = 0.006), and hunger (r = -0.180, p = 0.004). This indicates that older participants tended to engage more in physical activity but reported lower appetite and food-related responses. Gender showed a positive association with health literacy (r = 0.201, p = 0.001) and medication adherence (r = 0.113, p = 0.071), meaning that females had higher literacy and slightly better adherence than males. Education level was positively associated with health literacy (r = 0.143, p = 0.019), indicating that participants with higher educational attainment had better health literacy. Physical activity had a weak positive correlation with medication adherence (r = 0.035, p = 0.576), suggesting little relationship between physical activity and adherence. Participants with higher adherence were somewhat more health-literate (r = 0.126, p = 0.041) but reported lower enjoyment of food (r = -0.142, p = 0.023). The study showed that participants who ate more slowly tended to report higher quality of life, with positive association with slowness in eating (r = 0.124, p = 0.048). (Figure 1).

Fig. 1.

Heatmap correlation between MARS-5, HL, QoL, and baseline characteristics

Predictive modeling for MA and HL by MLP

The MLP model was applied to assess the relationship between MA, HL with other predictors. The model was built keeping MARS and HL as dependent/output variables, while other variables were considered input or independent variables. The two neural network models, based on two dependent variables, shared similar input layers (independent variables) consisting of 18 factors (e.g., age, gender, BMI, QoL, eating pattern, and smoking status) and utilized 171 units. In the hidden layer, Model-1 (HL) had 9 units, while Model-2 (MARS) had 15 units. The models were trained using a 70:30 split, with 70% of the data used for training and 30% reserved for testing. During training, the models achieved a cross-entropy error of 120.827 with an average incorrect prediction rate of 31.1%. Examining the categorical outcomes separately, 46.7% of health literacy cases and 15.6% of medical adherence cases were misclassified. The stopping rule was based on one consecutive step with no decrease in error, and training was completed rapidly in under one second (0.67 s). On the testing dataset, the model showed good generalization with a cross-entropy error of 14.292 and an average incorrect prediction rate of 21.4%. Specifically, 35.7% of HL cases and 7.1% of MARS cases were misclassified, indicating that the model performed reasonably well in predicting both HL and MARS outcomes. The “Excellent HL” category stood out with an AUC of 0.930, with strong model performance in identifying patients with excellent HL. Overall, the MARS model demonstrated balanced classification performance, while the HL model performed well across the different HL levels. (See Figs. 2 and 3). In the MARS model, biological factors dominated, with BMI (100%) and age (89.4%) as the main predictors, while eating-related variables showed moderate influence, with food responsiveness (48.8%), hunger (41.2%), and emotional undereating (43.2%) being most notable. In the MARS model, quality of life exhibited a normalized importance of 92%, highlighting its connection. For the HL model, hunger (75.5%), emotional overeating (72.6%), and satiety responsiveness (74.6%) stand out as key predictors. Both models agreed on the significance of emotional undereating (43.2% in MARS, 64.9% in HL), though the HL emphasizes behavioral factors like food fussiness (70%) and eating speed (54.1%), which gain greater importance compared to their values in the MARS model (21.2% and 37.4%). (See Fig. 4).

Fig. 2.

AUC curve for MARS-5 and HL classification

Fig. 3.

AUC curve for MARS-5 and HL classification

Fig. 4.

Normalized importance of the associated factors with MARS-5 and HL

Discussion

In this study, MLP modeling provided good predictive accuracy by modeling health literacy and Medication adherence determinants of old CMD Arab patients. This approach is recommended when traditional modelling cannot adequately address non-linearity and other assumptions. By partitioning the data into training and testing sets, the model demonstrated good performance in classification.

The study found low medication adherence rates among participants of the study, aligning with global trends in chronic disease patients, especially in LMICs [11–16]. Similar rates were reported in other studies [25, 26], with 60.7% non-adherence in Italy [27] and 51% in Sudan [28], while some reported better adherence [29]. Unlike prior literature highlighting exercise as a facilitator of adherence [25], this study found a positive but weak correlation between exercise frequency and adherence. Forgetfulness and intentional skipping were key barriers, reflecting broader issues of health engagement and treatment burden [8, 9].

BMI and age were major predictors of adherence, consistent with links to polypharmacy risk [27]. The positive correlation between HL and MA aligned with meta-analyses [18]. The study confirmed low health literacy among CMD patients in the Arab region, as supported by international studies [30–33]. Disparities in perceived well-being matched previous findings on QoL variability in chronic disease based on symptoms and self-management [37]. The study also noted a positive but weak correlation between HL and QOL, was observed supporting with the findings [38, 39].

The study also identified an association between psychological QoL and emotional overeating. This mirrors findings from [40]. Regarding patients specifically, 18% had Low QoL, 40.5% Moderate, and 41.5% High QoL, consistent with previous research [41, 42]. The ML model further indicated that BMI, quality of life, and eating behaviors, particularly tendencies related to emotional eating and hunger, are strong predictors of health literacy levels among patients. This supports existing literature on obesity’s bidirectional relationship with HL [34]. Furthermore, the study identified a positive association between HL and weekly physical activity among patients, a relationship supported by previous meta-analyses and systematic reviews [35, 36].

Practical implications

Based on the study’s results, which identified a very high rate of low medication adherence (82.4%) and key predictors included BMI, QoL, and specific eating behaviors, several specific interventions can be designed. To improve adherence, clinicians should create BMI-integrated counseling that directly links medication effects to patients’ weight management goals and employ QoL-focused messaging. To enhance health literacy, which was strongly predicted by hunger and emotional eating, education programs should use food-based analogies to explain complex health concepts. A specific learning program should be designed that guides patients into more intensive, visual-based support addressing differing relationships with eating behaviors, health roles, and improving quality of life.

Limitation

This study has limitations that need to be acknowledged. Its cross-sectional design prevents establishing causal relationships between predictors, while the convenience sampling from three Saudi hospitals may limit generalizability to the broader Arab population, and self-reported measures could introduce recall or social desirability bias. Additionally, unmeasured confounders (e.g., healthcare access, cultural beliefs) might influence outcomes. Despite these constraints, the study provides valuable insights using robust methodologies like validated instruments, advanced modelling technique MLP, and effect size, relevant to the target participants. Another limitation is a relatively modest sample size, which may affect the generalizability of the MLP model, which can be enhanced as larger datasets become available.

Conclusion

The study concluded that low MA and high levels of HL are common among CMD Arab patients. The factors, especially BMI, quality of life, physical activity, and emotional eating, are key contributors to these challenges and therefore, there is a need for the implementation of culturally relevant approaches that include health education, lifestyle changes, especially eating and physical activities manage CMDs more effectively.

Abbreviations

CMDs

Cardiometabolic diseases

HL

Health literacy

MA

Medication adherence

MARS-5

Medication adherence report scale

HLS-EU-Q16

European Health Literacy Questionnaire (16-item)

GLTEQ

Godin leisure-time exercise questionnaire

AEBQ

Adult eating behavior questionnaire

QoL

Quality of life

MLP

Multi-layer perceptron

BMI

Body mass index

AUC

Area under the curve

SDs

Standard deviations

SPSS

Statistical product and services solution

WHOQOL-BREF

World Health Organization Quality of Life – Brief Version

Author contributions

Hanan F. Alharbi: Conceptualization, Methodology, Data curation, Writing – original draft, Writing – review & editing. Abbas Al Mutair: Supervision, Project administration, Writing – review & editing, Funding acquisition. Muhammad Daniyal: Formal analysis, Methodology, Software, Writing – original draft, Writing – review & editing.

Funding

This work was supported by Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R441), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia.

Data availability

Data is not publically available and can be requested from corresponding authors.

Declarations

Ethics approval and consent to participate

The study adheres to ethical guidelines and was approved by the Institutional Review Board (IRB), King Abdullah Bin Abdulaziz University Hospital, Kingdom of Saudi Arabia, with the ethical no HA-01-R-104.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.