Dyslipidemia is a metabolic alteration characterized by elevated levels of cholesterol, triglycerides (TGs), and Low-Density Lipoprotein Cholesterol (LDL), as well as a decrease in High-Density Lipoprotein Cholesterol (HDL) levels. Worldwide, dyslipidemia presents as an exponential health problem with severe consequences and is considered one of the main risk factors for ischemic heart disease, cardiovascular disease, stroke, coronary heart disease, and type 2 diabetes mellitus (T2DM), which is the principal cause of death in adults in Mexico (1, 2). Pirillo et al. (3) have pointed out that ischemic heart disease reached a total of 3.78 million deaths in 2019, with high plasma LDL being the principal cause. These authors also reported between 0.61 and 2.73 million deaths due to ischemic stroke, a strongly associated condition. Similarly, there is a high variation in the number of deaths between countries, presumably due to regional differences and types of dyslipidemia. According to the same authors Pirillo et al. (3), low plasma HDL levels have been the most common type of dyslipidemia in Latin America since 2005, followed by hypertriglyceridemia and high plasma LDL levels.

According to the National Cholesterol Education Program Adult Treatment Panel III (ATP III) criteria, the classification of lipid profile dyslipidemias includes four types (see Table 1). Hypertriglyceridemia is a common lipid abnormality characterized by elevated triglyceride (TG) levels, often affecting individuals with visceral obesity, metabolic syndrome, and type 2 diabetes mellitus (T2DM) (4, 5). On the other hand, hypercholesterolemia is associated with high levels of LDL or CHOL and may also be present in individuals with a genetic disorder leading to elevated cholesterol levels (6). Hypoalphalipoproteinemia is frequently observed in people with coronary artery disease and is characterized by low levels of plasma high-density lipoproteins (HDL) (7). Finally, mixed hyperlipidemia, a genetic disorder involving higher cholesterol and triglyceride levels, contributes to the development of coronary artery disease.

In general terms, there are potential risk factors such as increased body mass index (BMI), an excessive dietary intake of saturated fat, and a sedentary lifestyle that contribute to developing a given type of dyslipidemia, a highly complex and heterogeneous set of conditions. This fact complicates the prognosis and diagnostics. In this regard, the widespread use of machine learning (ML) has allowed the application of computational intelligence tools as diagnostic tools for medical issues based on data acquired from analyzed patients. Therefore, such ML models (trained by medical guidance) have been successful in helping doctors to determine medical conditions with improved accuracy in a timely manner (8).

A study proposed by Cui et al. (9) uses ML to predict the risk of dyslipidemia in steelworkers by studying a set of standardized outcomes. They acquired the data by surveying anthropometric data, habits, personal status, and working details. Finally, they apply a Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) algorithm, showing excellent performance in predicting dyslipidemia in steel and iron industry employees.

Machine learning has emerged as a valuable tool in predicting dyslipidemia and related conditions based on patient data. For instance, Cui et al. (9) used a recurrent neural network and LSTM algorithm to predict dyslipidemia in steelworkers, achieving excellent accuracy. Lee et al. (10) correlated facial characteristics with hypertriglyceridemia using Naive Bayes classifiers, while Pina et al. (11) showed that a neural network outperformed the Dutch lipid score in predicting dyslipidemia in specialized lipid clinics.

Hatmal et al. (12) used ten ML techniques to predict dyslipidemia with an accuracy of 0.75, considering CD36 protein levels, lipid profile, blood sugar, gender, and age. Similarly, Kim et al. (13) classified and predicted overweight/obesity, dyslipidemia, hypertension, and T2DM using a deep neural network model based on nutritional intake data from Korean citizens. For each disease risk, the accuracies achieved were 0.62496, 0.58654, 0.79958, and 0.80896, respectively.

Dyslipidemia is a complex and heterogeneous condition with potential risk factors such as increased BMI, excessive dietary intake of saturated fat, and a sedentary lifestyle. In this context, machine learning models trained on medical data have shown promising results in improving the diagnosis and prognosis of this condition.

However, recent research indicates that the impact of dyslipidemia on cardiovascular health can vary between men and women due to hormonal, genetic, and lifestyle differences. By analyzing gender-specific differences in dyslipidemia, we can identify unique risk profiles, treatment responses, and underlying mechanisms that may contribute to cardiovascular outcomes. Tailoring interventions based on gender-specific dyslipidemia patterns can lead to more targeted and effective therapies, ultimately improving cardiovascular health for both men and women. This approach also highlights the importance of recognizing and addressing gender-related disparities in dyslipidemia management to optimize patient outcomes and reduce the burden of cardiovascular diseases.

Historically, clinical trials are predominantly done in men, excluding women, even in studies with cells and mice (only male). A review studies the significant causes of diseases by bias in sex and gender. The authors express the influence of differences between sex and gender in genetics, implying affection in diagnosing and treating illnesses (14).

In this context, the present work provides a machine-learning approach to characterize the particularities of men and women with a given type of dyslipidemia (hypertriglyceridemia, hypercholesterolemia, hypoalphalipoproteinemia, as well as mixed hyperlipidemias), identifying the association with clinical factors, biochemical screening, family health history, dietary information, and additional risk factors in order to provide features that can be monitored by health authorities to decrease the risk of long-term complications caused by lipid abnormalities in the study population. While dyslipidemia is a significant risk factor for serious diseases, we acknowledge that our analysis does not incorporate a specific time frame within which an individual might develop the disease. Instead, our study aims to elucidate the underlying risk factors associated with dyslipidemia, providing valuable insights into its etiology and contributing factors. The criteria used in this study to classify dyslipidemia types are shown in Table 1.

We used a 32 GB RAM, 3.50 GHz, Intel Xeon® Dell® Workstation to perform all calculations. R v. 3.6.1 with RStudio and Python v. 3.10.7 were used as programming languages. Purposely, these resources are readily available for implementation in most hospital informatics settings.

The problem of abnormal TG levels can develop based on different factors influencing individuals depending on their lifestyle. Moreover, LDL levels tend to be higher in men than in women until menopause. Hence, in this study, we initially separated the data by gender to obtain the most crucial care features according to the type of dyslipidemia. To identify the potential features by gender and type of dyslipidemia, we applied SMOTE as a resampling method due to class imbalance and three machine learning algorithms, namely VIM of RF, XGBoost, and GBM.

Once we obtained the results from the aforementioned algorithms, we considered displaying at least the top ten most important variables (ranked) that influence each type of dyslipidemia. Each result table shows a different subset of features for each gender and type of dyslipidemia by applying VIM of RF, XGBoost, and GBM.

The results obtained for hypertriglyceridemia are presented in Table 2, followed by the results for hypercholesterolemia in Table 3, as well as the essential variables for hypoalphalipoproteinemia, displayed in Table 4, and finally, the results for mixed hyperlipidemias in Table 5. Summarized general data from the total cohort is presented in Supplementary Table 4.

Subsequently, each algorithm generated subsets of variables, which were used to select the best features. To perform this feature selection process, we applied RF, which was optimized by grid search method (27) (resulting in varying mtry and ntree values for each gender and dyslipidemia type). We employed 10-fold cross-validation with ten repeats to evaluate the performance. Following this, we conducted 30 independent executions with different seeds to ensure robustness and approximate a normal distribution. This approach aligns with similar practices observed in relevant studies (28, 29). The evaluation was based on balanced accuracy, serving as the primary criterion for assessment.

To measure the performance of the RF model, the metrics B.ACC, SENS, and SPC were considered; likewise, it was necessary to apply SMOTE due to the unbalanced dataset. Table 6 shows each result of RF by using the different subset of variables obtained by VIM of RF, XGBoost, RPART, and SHAP, for men and women, as well as the respective parameter tuning and standard deviation (SD).

In the case of men with hypertriglyceridemia, the subset of features obtained by XGBoost achieved the best RF performance with a B.ACC of 82.77% and SD of 1.26. The top variables of this subset showed the influence of overweight, where the first three variables are related to it and body mass index (BMI), followed by age, sleep disturbance (SLPD4) and FYI (SLPSNR1), anxiety as a trait (TR.ANX), smoking practice (SMOKE), somnolence (SLP3), alcohol consumption (ALCOHOL), soy oil consumption (SOYAOIL), glucose levels (FPG), and medical history of the mother with hypertension (M.HPT).

Moreover, for women, the subset of variables obtained by VIM of RF achieved the best performance with a B.ACC of 82.50 and an SD of 1.08, where the principal variable was uric acid levels (URIC) [several studies (30, 31) have found an association between high uric acid and hypertriglyceridemia]. The other variables in this subset include glucose levels (FPG), body mass index (BMI), Systolic blood pressure (SBP), weight, age, Diastolic blood pressure (DBP), Waist circumference (WAIST), sleep disturbance (SLPD4), somnolence (SLP3), height, snoring (SLPSNR1), and smoking practice (SMOKE). All these variables are considered risk factors contributing to the development of hypertriglyceridemia (32, 33).

For hypercholesterolemia, the variables obtained by GBM achieved the best RF performance for men, with a B.ACC of 83.69% and an SD of 1.52. The principal variables found by this model denote a close relation between being overweight as represented by (WEIGHT, WC, AGE, and HEIGHT), sleep disturbances (SLPSNR1, SLP3, and SLPSOB1), anxiety disorders (TR.ANX), and habits such as consumption of flavored soda (FLAVSODA) and smoking (SMOKE).

In the case of women with hypercholesterolemia, the best performance was obtained by the subset generated by VIM or RF with a B.ACC of 79.74% and SD of 1.16, where the anxiety disorders (TR.ANX) and uric acid levels (URIC) were the principal variables, as well as frequently consuming some foods like chicken liver (LIVERSTK), bread (WHBREADSL), oatmeal bowl (OATMEAL), and margarine (MARGARIN), likewise, variables related to sleep disorders like the time to fall asleep (TIM.SLP) and sleep short duration (SLPOP1), followed by low physical activity (METS.low), smoking and history of obese parents (P.OBS).

For men with hypoalphalipoproteinemia, the best subset of variables was presented by GBM with a B.ACC of 80.50% and SD of 1,29, being variables related to overweight the best qualified (WEIGHT, WC, and HEIGHT), as well as age, followed by indicators of sleep disorders like sleep disturbance (SLPD4), snoring (SLPSNR1) and somnolence (SLP3). Likewise, habits of alcohol consumption and smoking, anxiety disorder, cream cheese consumption (CRMCHSPOO) and elevated uric acid levels (URIC).

Similarly, the VIM of RF was the best subset of variables for women with hypoalphalipoproteinemia, with a B.ACC of 83.65% and SD of 1.22. In this case, the principal variable was elevated uric acid levels, followed by snoring (SLPSNR1) and variables closely related to overweight (BMI, WC, WEIGHT), as well as glucose levels (FPG) and blood pressure levels (SBP and DBP) denoted their presence as risk factors, finishing with the consumption of alcohol and bread (WHBREADSL), as well as anxiety.

Finally, the subset of variables obtained by VIM of RF got the best performance for men with mixed hyperlipidemia. In this case, the main variables were closely related to food consumption such as atole without milk (OATMEAL1), oatmeal bowl (OATMEAL), a teaspoon of margarine (MARGARIN), a glass of table wine (TABLEWIN), safflower oil (SAFFLOWR), rum, brandy or tequila (HARDLQUR), zapote (FREQ025), a tablespoon of cream cheese (FREQ005) and olive oil (OLIVEOIL), as well as, ALCOHOL, BMI, history of a parent with dyslipidaemia (P.DSLP), and age.

For women with mixed hyperlipidemia, the variables obtained by VIM of RF with the best-ranked factors were BMI, age, and snoring, followed by glucose levels, waist circumference, sleep short duration, uric acid levels, smoking, height, anxiety, alcohol consumption, Systolic Blood Pressure, and a glass of flavored sugar water (SUGRDRNK).

In what follows, we will discuss the present analysis's expected and novel findings to contextualize the potential value of public health interventions.

In order to determine the significance of studying males and females separately, a significance analysis was conducted using the chi-squared test. The results indicated a strong relationship between gender and the prediction of dyslipidemia types and their critical factors.

The significant associations found for SEX in all dyslipidemias type further emphasize the importance of gender as a significant factor influencing dyslipidemia prediction. Therefore, conducting separate analyses for males and females was crucial to gain a comprehensive understanding of the underlying factors associated with dyslipidemia in each gender group. The results of this significance analysis can be seen in the Supplementary Tables 5–8.

In the case of men with hypertriglyceridemia, several known associations arise. That is the case of overweight (34–37), age (38, 39), and waist circumference (40, 41). Additionally, we discovered a set of relatively new yet significant predictors whose relevance and mechanisms concerning hypertriglyceridemia in men are still to be determined, such as anxiety, tomato sauce consumption, and history of hypertension in the mother. Regarding the association between anxiety and hypertriglyceridemia, van Reedt Dortland and collaborators have identified a potential role of tricyclic antidepressant drugs (42). In contrast, other authors have identified an increased risk of hypertriglyceridemia in patients with psychiatric diseases without relation to specific pharmacological treatment (43).

The case of tomato sauce consumption presents some contradictory features. At the same time, some authors have described a protective role of processed tomato products to post-prandial oxidation and inflammation (both associated with dyslipidemias) in healthy weight subjects (44–46). In contrast, others have related processed foods (including tomato sauce) to hypertriglyceridemia (47, 48).

Since these studies differ in the methods and types of populations under investigation, differences may be explained by such disparate approaches. Hence definite associations need to be further studied with properly defined research methods.

No previous studies have directly linked maternal hypertension history to hypertriglyceridemia. Interestingly however, is the fact that there is an unusual prevalence of hypertriglyceridemia in small populations with known risk factors for pregnancy-associated high blood pressure (49–52), though, at this stage, an actual association is still to be further validated in more extensive population studies.

Similarly, in the case of women with hypertriglyceridemia, the best predictors were some known factors such as AIP (a prominent feature by construction) as well as BMI, age, and cola drink consumption. Other metabolic features appear, such as glucose and uric acid levels and also raw tomato consumption. Regarding the role of high fasting glucose levels in the presence of hypertriglyceridemia, reports have long been made, particularly by driving mechanisms of endogenous hypertriglyceridemia (35, 53, 54). The fact that FPG is a better predictor for hypertriglyceridemia in women than in men may be related to the effects of hormone (in particular, estrogen) metabolism in lipid and glucose processing biochemical pathways (55–57).

Elevated uric acid levels have been previously associated with hypertriglyceridemia, both in extensive cohort studies (31, 52, 58–60), population-based research (61–64), and biochemically-based analyses (65–69). Unlike processed tomato products, whose effects on hypertriglyceridemia are ambiguous (as previously discussed), raw tomato consumption has been acknowledged as a protective factor (44) against dyslipidemia in general and hypertriglyceridemia, in particular, (70–73).

Regarding men with hypercholesterolemia, some of the main predictors are (unsurprisingly) meat-based products with high lipid contents such as tacos al pastor (shepherd style), carnitas, and longaniza (74–76). There is evidence that consuming fatty meats, such as beef, pork, and lamb, may contribute to the development of hypercholesterolemia.

For instance, one study published in the American Journal of Clinical Nutrition found that a diet high in saturated fat, such as that found in fatty meats, was associated with an increase in LDL cholesterol that can, in turn, contribute to the development of cardiovascular disease (77). Another study published in the American Journal of Epidemiology found that individuals who consumed a diet high in red and processed meats had a higher risk of developing hypercholesterolemia than those who consumed a diet low in these types (78).

Aside from fatty meat products, other predictors are foods such as chocolate powder, cream cheese and anthropometrics such as weight and height (79, 80). Some evidence, for instance, suggests that chocolate consumption may be associated with a modest reduction in cholesterol levels, although the effect may be small and may depend on the type of chocolate and the individual.

Several studies published in the American Journal of Clinical Nutrition and the European Journal of Clinical Nutrition found that cocoa and chocolate intake was associated with a slight reduction in total cholesterol and low-density lipoprotein (LDL) cholesterol and that the effect of chocolate on cholesterol levels may be influenced by the type of chocolate consumed, with some studies suggesting that dark chocolate may have a more significant effect on cholesterol levels than milk chocolate (81–83). In contrast, another study recalls that these effects may come via activating flavonoid metabolism and anti-oxidant pathways (84).

In the case of women with hypercholesterolemia, there are well-known factors such as age (85, 86), pork rind (87, 88), mayonnaise consumption (89, 90), and BMI (91, 92). Other less-known predictors emerge from our study. Such is the case of sleep disturbance. Abnormal sleep conditions are gradually being recognized as relevant players in metabolic and cardiovascular diseases (93–95). However, it is noteworthy that most studies relating hypercholesterolemia with sleep disturbances center on the possible effects on sleep induced by drugs such as Pravastatin and Lovastatin (96–100).

The main predictors found for alphalipoproteinemia in men were waist circumference and BMI (101, 102), as well as conditions such as anxiety (103, 104), and consumption of seafood (105, 106) and plums (107). In contrast, in women, selected features were known metabolic state and anthropometric markers such as AIP (108), glucose levels (109, 110), BMI and waist circumference (101, 102), also uric acid levels (111, 112); consumption of high fat or high caloric foods like pork meat, flavored soda, Oaxaca cheese, and bacon (113). Interestingly snoring while sleeping was also a relevant predictor for alphalipoproteinemia in women. Though a direct association of snoring with female alphalipoproteinemia has not been reported, a population-based study has indeed associated self-reported snoring with dyslipidemia, high total cholesterol, and high low-density lipoprotein cholesterol in obese individuals in rural China (114).

Mixed hyperlipidemias in men were best predicted by: AIP (115, 116), waist circumference, BMI (117, 118), age (119), as well as dried chile peppers consumption (DRYCHILES) (120–122), as well as drinking whole milk (MILKGLASS) (123, 124), alcohol (125–127), sweet bread (SWEETBRD) (128, 129), and orange (ORANGE) intake (130–132). In the case of women with mixed hyperlipidemias top predictive features were: BMI (117, 118), age (119), snoring (114), glucose levels (133, 134), uric acid levels (64), smoking (133), and anxiety (135), but also alcohol (125, 136) and flavored sugar water (BACONSLC) (137, 138) consumption.

By focusing on identifying risk factors without a time frame, our study lays the foundation for future investigations that could incorporate temporal aspects for predicting the onset of dyslipidemia or subsequent development of CVD. The findings from our research can serve as a basis for developing predictive models that integrate time-based parameters, enabling more accurate and clinically relevant disease prognosis and management.

In this work, the application of machine learning models in a cohort of Mexico City allowed the identification of subsets of attributes acting as risk factors associated with several types of dyslipidemias. Multi-feature diagnostics, i.e., the diagnosis based on different aspects, is considered essential to support healthcare providers as it allows early detection of patients at the most significant risk of developing a type of dyslipidemia, which supports the development of strategies for prevention, treatment, and prognosis the condition.

The separation by gender allowed the discovery of differences between subsets of risk factors associated with each type of dyslipidemia.

Even when we obtained high-performance models with this particular data and the support of SMOTE, it is possible to note that the best classifiers identified risk factors in men with hypercholesterolemia (with a B.ACC of 83.69%) and women with hypoalphalipoproteinemia (with a B.ACC of 83.65%). Therefore, the exploration of other ML models and the continuous update of the data set may not be ruled out in future work to improve the values of the metrics and predict the development of dyslipidemia types.

The original contributions presented in the study are included in the article/Supplementary material, further inquiries can be directed to the corresponding author/s.

The studies involving humans were approved by Research Ethics Board for Biomedical Research in Humans by the National Institute of Cardiology Ignacio Chavez-Protocol approved with key 13-802. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants' legal guardians/next of kin in accordance with the national legislation and institutional requirements.

GG-E designed computational strategy, implemented computing code and algorithmics, evaluated performance measures, co-supervised the project, and drafted the manuscript. TRP-Z guided the clinical approach and provided feedback to the modeling. TR-d implemented computing code and algorithmics and contributed to drafting the manuscript. MM-G performed clinical, sociomedical, and health policy research contributed to drafting the manuscript. LG-M supported data curation. MFM-M performed a clinical assessment. LMA-G contributed to clinical assessment. LMA-G, TRP-Z, and GV-A reviewed clinical results. EH-L devised the overall study strategy, co-supervised the project, performed the technical assessment, and revised and edited the manuscript. All authors read and approved the submitted version of the manuscript.