Patients are becoming more educated regarding their health care and studies have shown that both diet and ethnicity with genetic variants influence cardiovascular disease (CVD) development (1-4). Since CVD is considered the leading cause of mortality worldwide, preventing, reducing and lessening the development of the disease is critical for improving mortality and morbidity rates (5). Lifestyle and daily diet vary considerably between countries and continents, and this has been shown to affect the development of CVD (6-11). Patients who are aware of this wish to know what effect they will have from a certain treatment, medicine or lifestyle change. From a medical professional point of view, it is crucial to have information on whether a recommended treatment or lifestyle change will be effective for an individual patient. Variations in the effects of drugs between patients have been observed and even more will be discovered in the future (12,13). From a health-care planning perspective, finances can be distributed more accurately there is available knowledge on whether a patient will experience the desired effect of a drug considered for prescription. Personalized medicine is currently being more commonly requested by patients, medical professionals and politicians.

Garlic (Allium sativum L.) has played a paramount role in traditional medicine, dating back to the ancient Egyptians. Garlic extracts have been shown to increase nitric oxide synthase activity, a critical part of vascular function and angiogenesis (14). Aged garlic extract (AGE), a garlic supplement, contains a number of active ingredients, one of these being S-allylcysteine (SAC) (15). SAC has antioxidant properties, it prevents lipid and protein oxidation, and is the most extensively studied component of AGE (16-18). Ahmadi et al (19) demonstrated that AGE was associated with a reduction in the levels of an inflammatory biomarker, a lack of progression of coronary artery calcification (CAC) and with increased vascular function. In their studies, Budoff et al (20-22) also demonstrated an association of AGE with a favorable improvement in oxidative biomarkers, vascular function and the reduced progression of atherosclerosis measured by CAC progression. Steiner et al (23) revealed that AGE exerted a beneficial effect on lipid profiles and blood pressure. The authors have previously demonstrated an association between AGE and improved microcirculation, a decrease in blood pressure and a positive effect on inflammatory biomarkers (24-26).

A method for predicting the outcome of various drugs and medical treatments is known as data mining. It is performed by analyzing large blocks of information and trying to find patterns, similarities or trends. Translating a large set of data to a trend or a new-found pattern can make the data easier to interpret and lead to new and useful information. The method is based on combining visualization of diagrams and different algorithms to finding patterns or trends. There are two starting positions, either driven by subject-specific information and hypotheses using data mining to confirm the theory or using data mining to generate data-driven hypotheses. These two approaches should be combined to avoid flawed conclusions. Probably the most well-known approach is the cross-industry standard process for data mining (CRISP-DM) (27-29). Using this method and approach, the authors previously developed an algorithm for predicting which patient will have a more favorable effect on CAC scores and blood pressure following the intake of AGE capsules (25). The AGE algorithm was developed using a European cohort. The aim of the present study was to validate the method of constructing an algorithm for prediction of which patient will have a more favorable effect of AGE on CAC and systolic blood pressure (SBP) using data accumulated in the USA.

In a previously published article by the authors, we developed the AGE algorithm on 46 patients in the Swedish cohort. To validate the method of developing the AGE algorithm in the present study, data from four studies were retrieved, performed in the USA, termed the USA cohort. The USA cohort consisted of a total of 78 patients who received AGE who were included in the present study. Following the aforementioned developmental steps of the prediction model, the model was validated using LOOCV. For each of the 78 subjects in the dataset, the authors computed the estimated probability of the i:th subject being in the 50th percentile with the optimal progression following 12 months of AGE supplement treatment. i:th represents the subject who is being tested on among the N observations. The following mean modeling key performance indicators were provided by the LOOCV: CAC: Accuracy, 65%; precision, 64%; recall, 69%; and SBP: Accuracy, 65%; precision, 66%; recall, 64% (Table V). In total, 12 features were selected to be in the model for sufficient accuracy, but to reduce overfitting. For CAC, the 12 features were the following: LDL, sex x cholesterol, sex x SBP, age x triglycerides (TG), BMI x LDL, CAC2, CAC x homocysteine, cholesterol x LDL, cholesterol x TG, Il-6 x SBP, IL-6 x SBP, IL-6 x TG and LDL2. An overview and each feature of the associated coefficient are provided in Table II. For SBP the 12 features were the following: Sex x DBP, sex x HDL, age x IL-6, age x CRP, CAC x IL-6, CAC x CRP, DBP2, HDL x SBP, HDL x CRP, homocysteine x IL-6, LDL x CRP and CRP2. An overview and each feature of the associated coefficient are provided in Table III.

The terms ‘personalized medicine’, ‘precision medicine’ and ‘stratified medicine’ are all used to describe the same concept. The majority of articles, reports and researchers use these terms interchangeably. The concept is to customize and tailor medical treatments, drugs and recommendations to each individual patient depending on the predicted response or risk of disease. Personalized medicine often uses diagnostic testing with molecular or cellular analysis to allocate patients to groups. In one aspect of personalized medicine, pharmacogenomics or pharmacogenetics, the genetic information of an individual is used to tailor medical treatments. Although this may appear as a very modern approach, medical professionals have been making decisions such as these for decades. It is well established that a number of diseases have a large hereditary component and patients are questioned about this for the purpose of determining whether they have a smaller or larger risk of developing a certain disease or condition. The medical professional then reaches a decision based on this information, prescribing the recommended medicine as ‘one-size-fits-all’, meaning that the majority of patients are then prescribed the same drug and dose. The decision behind this is that the majority of studies are based on broad population averages. However, not many patients fit into this description and numerous drugs will not function in the same manner for everyone. The majority of drugs are also prescribed along the basis of ‘trial and error’. Treatment begins with the most common and popular drug and then in the case that the patient develops side-effects or does not respond in as expected, the drug is changed to the second drug of choice on the list for that condition. This often results in frustrated patients and tired doctors. The notion behind personalized medicine is to avoid this waste of time and money by selecting the right medicine or recommendation to start with.

The present study used this notion to create an algorithm that could be applied to patients who wish to know whether AGE will have a positive effect on their health. As described previously, patients with diverse lifestyles, certain genes and diets respond differently to drugs and dietary supplements (1-4,6-11). The authors have previously studied patients who are at an intermediate-to-high risk of developing cardiovascular events and created an algorithm, the AGE algorithm, to predict the effects of AGE on CAC progression and blood pressure of an individual (25). The algorithm is based on the CAC score, blood lipids and blood pressure and could predict, with 80% precision, which patient will have a significantly reduced CAC progression after 12 months of AGE supplement (25). The same type of algorithm, the one developed in the present study, could predict with a 64% precision score, which patients will have a significantly reduced CAC progression after 12 months of AGE supplementation. The algorithm could predict with a 66% precision score which patients will have a significant reduction in blood pressure after 12 months of AGE supplementation. Unlike existing cardiovascular risk calculators, such as the Framingham Risk Score (https://www.mdcalc.com/calc/38/framingham-risk-score-hard-coronary-heart-disease) or the ASCVD Risk Estimator (https://www.mdcalc.com/calc/3398/ascvd-atherosclerotic-cardiovascular-disease-2013-risk-calculator-aha-acc), which predict the likelihood of future cardiovascular events based on broad population data, the model used herein is specifically designed to predict individual patient response to a defined intervention, AGE supplementation. While previous studies have demonstrated average cohort-level effects of AGE, to the best of our knowledge, no previous study to date has developed or validated a personalized prediction tool estimating individual treatment responses (33,34). By applying data science methodologies incorporating interaction terms and non-linear relationships, this model represents a novel, practical step toward personalized, supplement-based cardiovascular prevention. The algorithm demonstrated a precision of 64% for CAC and 66% for SBP, which is notable given the modest sample size. To the best of our knowledge, this is the first externally validated prediction tool tailored to identify which patients are most likely to benefit from AGE, providing a novel contribution to precision prevention strategies in cardiovascular care. It is clear there is a potential in predicting which patients would benefit the most from AGE supplement treatment using a limited sample (25). Therefore, further validation of the method of developing the algorithm on another cohort is required to broaden the concept of using it in day-to-day medical consultations with patients interested in dietary supplements.

When developing the model, a combination of variables and engineered features was made available for the model to use. A balance between model complexity and performance was sought to maintain the number of features to a minimum, to avoid overfitting and increase model stability. Initially, a model with 10% of the available features was tested, i.e., 10 out of 105 features in total. The model was performing well; however, a model with an additional two features was tested, which was found to outperform the model with 10 features. As a result, a model with 12 features was selected as the final model. Logistic regression was selected for its interpretability and suitability for binary outcomes in clinical applications. While other machine learning methods could be used, logistic regression provided a practical balance of transparency and predictive performance, particularly in a limited sample size.

Overall, the models generalized well on the test data. When selecting the validation method, 10-fold cross validation was explored; however, it was deemed to be too sensitive to the random sampling effect when creating the folds. It is important to realize that in a 10-fold cross validation, 10% of the sample is used as test, while only having 78 patients, the test fold consists of only 8 patients. With a limited sample size, the probability of a test set being skewed is high. LOOCV is a reliable method that produces an unbiased estimate of the performance of a model, and it is recommended that it is used in situations where data are limited. Although LOOCV is computationally expensive, the limited number of observations render the computational cost negligible. Therefore, with limited data, the random sampling effects using 10-fold cross validation, limiting bias and enabling testing on unseen data renders it the optimal validation method to use.

As aforementioned, the prediction model performs better on unseen data, i.e., the more data it is used on. Generally, the effect from the amount of study data available can be broken down into two parts. First, the negative random effect on the partitioning of data into training sets and test sets diminishes the more data is available, i.e., the representativity of both the training and test set increases. Second, more data allow for the addition of more categorical predictor features to the modeling phase. Increased clinical information, such as lifestyle, certain genes and dietary intake in categorical form would increase the performance of the model. This is due to the fact that a model is built on a sample with a higher variance, i.e., a larger sample, which results in a model with an intrinsic ability to generalize more effectively on unseen data. Overall, the present study indicates the feasibility of developing predictive models classifying patients into who would have the best results from the AGE supplement or not. This predictive modeling framework could also be applied to other areas of biomedical research. Recent advances in AI-driven treatment prediction, organ-on-a-chip platforms for drug screening, engineered cardiac organoids for cardiovascular disease modeling and artificial vascular graft technologies illustrate the expanding opportunities for integrating predictive models into precision medicine and translational cardiovascular research (35-38).

It is important to note that baseline differences between the Swedish and USA cohorts, particularly in age, LDL and SBP, could influence model performance across populations. Although the model incorporated interaction terms and non-linear features to account for these variations, further validation in larger, more diverse cohorts will be necessary to confirm broader applicability. The observed decline in model performance when applied to the USA cohort is likely explained by differences in baseline variable distributions, particularly in age, LDL, SBP and CAC scores, which can alter predictor-outcome relationships. Additionally, as feature selection was based on the Swedish cohort, the relative importance of certain variables and interactions may have differed in the USA population, contributing to reduced accuracy and precision. These findings highlight the importance of accounting for population-specific variability in model transportability and support the need for broader validation in diverse clinical populations.

The requirements of gathering data to assess who would benefit the most from AGE supplement treatment are not high: ordinary blood tests and a CT scan can be utilized to gather the necessary variables. However, further studies, tests and modeling are necessary before a model can be used to assess which patient will have a favorable effect of AGE supplement treatment or not.

The present study had certain limitations which should be mentioned. The data in the present study were acquired from four studies performed previously. Given the variation in collected variables and the limited number of patients, these could potentially affect the scores. Furthermore, the relatively small sample size, with 46 patients in the initial model and 78 in the validation cohort, may limit the power of the model and increase the risk of overfitting despite the use of recursive feature elimination and leave-one-out cross-validation. Thus, larger, prospective studies are warranted to further validate and improve the predictive accuracy and generalizability of the model. A larger study cohort would probably reduce the random effect of within-group variation. CAC progression was defined as the absolute change in CAC. A potential site variation in regard to different brands of the CT scan itself and the analyzing software might also inflict some variations. The interpreting radiologist differed, which could be observed as a variance of weakness, but also a strength. The serum level of SAC was not measured to ensure patient compliance in the studies. All patients were advised not to change their lifestyle and diet apart from taking the AGE capsules; however, studies have indicated that patients enrolled in studies are more likely to adopt lifestyle changes (39,40). The lifestyle changes were not assessed formally to evaluate the effect on the endpoint. A randomized trial of diet and lifestyle would be needed to better assess these potential interventions.

In conclusion, the present study demonstrates that it is possible and realistic to develop predictive models, classifying patients into who would have the optimal results from AGE supplement treatment or not. The algorithm was able to predict with 64% precision which patient would have a significant reduction of CAC progression using the AGE supplement. With the same type of model, we could also predict with 66% precision which patient would have a significant blood pressure-lowering effect using the AGE supplement.