Introduction

Cardiovascular disease (CVD), including coronary heart disease, cerebrovascular disease, and peripheral arterial disease, remains a significant global health challenge. In 2021, CVD affected over 500 million people and caused 20.5 million deaths, accounting for nearly one-third of global mortality [1]. Despite a decline in age-standardized mortality rates, regions such as the Western Pacific, including Malaysia, have witnessed a 47.4% increase in CVD-related deaths, highlighting the persistent burden of these conditions [2]. CVD continues to be the leading cause of non-communicable disease (NCD) deaths, responsible for 44% of the 41 million NCD-related deaths globally in 2019 [3,4].

In response to the growing burden of CVD, the “Regional Action Framework for Noncommunicable Disease Prevention and Control in the Western Pacific” emphasizes the importance of targeted prevention efforts, particularly through screening high-risk individuals. Various risk assessment tools, such as the Framingham Risk Score (FRS), Pooled Cohort Equations (PCE), and WHO/ISH risk charts, have been developed to estimate CVD risk [5–8]. Recent advancements in risk prediction have led to the development of updated tools like SCORE2 and the Revised Pooled Cohort Equations (RPCE), which integrate contemporary population data and refined methodologies [9,10]. However, these tools require thorough external validation before being applied to diverse populations. For example, the WHO CVD risk score, while demonstrating good discriminatory power in a Chinese population, was found to overestimate CVD events, highlighting the necessity for local validation and calibration [11].

Despite these advancements, there remains a critical gap in the availability of an updated CVD risk score specifically tailored for the Asian population. This gap forces clinicians in Asian countries to rely on risk models predominantly developed for Caucasian populations, leading to potential inaccuracies and inconsistent risk stratification, particularly in regions like Malaysia [12]. The development and validation of population-specific risk models are therefore imperative to improving CVD risk prediction and ensuring effective prevention strategies across diverse global populations.

Machine learning (ML) offers a robust alternative to traditional risk-scoring systems by encompassing a range of statistical techniques and algorithms that allow computers to learn from data and improve decision-making without explicit programming. ML algorithms such as Logistic Regression (LR), Support Vector Machines (SVM), Random Forests (RF), Extreme Gradient Boosting (XGB), Naïve Bayes (NB) and Neural Network (NN) have shown significant potential in healthcare applications [13–16]. Unlike traditional statistical methods that often rely on predefined assumptions and models, ML algorithms can adapt and learn from data without stringent assumptions, uncovering hidden patterns and relationships. The implementation of feature selection methods in ML further enhances overall model accuracy [17].

Studies have demonstrated that ML models outperform conventional CVD risk scores like the FRS across various populations. In the United States, a study showed that models like the multi-layer perceptron (MLP) provided better predictive accuracy than traditional methods [18]. Similarly, in the UK, the researcher found that advanced ML models, including Gradient Boosting and AutoPrognosis, outperformed traditional risk scores in predicting CVD events [19]. In Australia, the researcher reported that ML models, such as LR, were more accurate to than the FRS in predicting long-term CVD mortality [20]. Similar trends were observed in Japan [21], China [22] and Korea [23], where ML models consistently outperformed conventional risk scores. These findings suggest that ML-based approaches offer more accurate and personalized risk assessments, potentially leading to improved clinical outcomes.

While ML has demonstrated its ability to outperform conventional CVD risk scores, its full potential in CVD risk assessment remains underexplored, particularly with advanced techniques like stacked ensemble learning (EL) and SHAP (SHapley Additive exPlanations) analysis. Stacked EL is a sophisticated ML approach that combines predictions from multiple models to create a more robust and accurate final model. This technique has proven highly effective in CVD mortality prediction, often surpassing the performance of individual ML algorithms and significantly enhancing patient outcomes [24]. Similarly, SHAP analysis, a powerful interpretability tool, has been employed to provide detailed insights into the contribution of each feature within complex ML models, thereby increasing transparency and trust in ML-driven healthcare applications.

Despite the success of these advanced ML techniques in CVD mortality prediction, their application in CVD risk assessment is still limited. This presents a critical opportunity to leverage these methods to develop models that are not only highly accurate but also interpretable—both of which are crucial in clinical settings. To address this gap, our study will utilize a dataset from Malaysia, a multiethnic country representative of the broader Asian population. By applying feature selection, stacked EL, and SHAP analysis, we aim to develop a CVD risk assessment model that is both precise and interpretable, tailored specifically to the Asian population, and capable of improving risk stratification and clinical decision-making.

Methods

Study design and setting

The objective of this study was to address the limitations of current CVD risk scores, which were primarily derived from Caucasian populations, by constructing a model tailored for the Malaysian population. We developed a CVD risk prediction model using data from the Responding to Increasing Cardiovascular Disease Prevalence (REDISCOVER) Study, which is a longitudinal community-based cohort initiated in 2007. This registry contains detailed demographic, clinical, and lifestyle information from a wide range of Malaysian individuals with no prior history of CVD aged 18 and older, recruited between 2007 and 2017. The primary outcomes assessed were fatal and non-fatal CVD events occurring within 10 years after baseline recruitment.

Participants underwent detailed assessments, and 30 key predictors were selected from an initial set of 202 features. The selected features were applied to construct and evaluate by employing both base and ensemble ML models. The performance of these models was compared against conventional risk scores like the FRS and RPCE to determine the most effective approach for CVD risk prediction in Malaysia. The study’s workflow and methods are shown in Fig 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Research workflow and methodology applied in this study.

https://doi.org/10.1371/journal.pone.0323949.g001

Results

Patient characteristics

Table 1 presents the baseline characteristics of the 5,688 participants included in this study. This table provides a univariate analysis of all variables, showing their distribution across the two groups (CVD event vs. no CVD event) and statistical significance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Comprehensive baseline characteristics and outcome of the complete dataset.

https://doi.org/10.1371/journal.pone.0323949.t001

Among these, 202 (3.55%) had a CVD incident within 10 years of follow-up. The mean age of participants who had a CVD incident was 58.28 (SD 9.9) years, while those who did not have a CVD event were slightly younger, with a mean age of 51.59 (SD 10.85). 64.01% of the patients who had CVD events were male, compared to 42.02% in the non-CVD group. The majority of those who experienced a CVD event lived in rural areas (56.44%) and were current smokers (21.29%), compared to 44.24% and 11.28% in the non-CVD group, respectively. Education levels differed significantly, with 37.62% of CVD event participants having a primary education compared to 23.42% in the non-CVD group. Clinical measures revealed a higher mean BMI (26.66 vs. 25.95), systolic blood pressure (149.00 vs. 134.56 mmHg), and diastolic blood pressure (84.55 vs. 80.10 mmHg) in the CVD group.

In terms of metabolic health, individuals who got CVD had higher glucose levels (6.59 vs. 5.50 mmol/L), higher triglycerides (1.97 vs. 1.68 mmol/L), and lower HDL cholesterol (1.14 vs. 1.25 mmol/L). The CVD group also had a greater incidence of diabetes (24.26% vs. 10.46%) and hypertension (41.58% vs. 23.42%).

Menopause frequency and medication use are presented in S1 Table for subgroups based on diabetes, high cholesterol, and hypertension, further broken down by the occurrence of CVD events. Women with CVD were more likely to be postmenopausal (80.25% vs. 47.15% overall). Across all three conditions, medication use was relatively high: 75.4% of patients with diabetes, 65.4% of patients with high cholesterol, and 73.7% of patients with hypertension were on medication. In the CVD group, medication use was slightly lower for diabetes, but more prevalent for hypertension and high cholesterol.

In conclusion, those who had CVD events were typically older men with higher BMI, blood pressure, glucose, and triglyceride levels. They also had a greater rate of smoking, diabetes, hypertension, and lower HDL cholesterol levels, emphasising the multifaceted nature of cardiovascular disease risk. There were significant differences in gender, age, geographical area, education level, smoking, systolic blood pressure, diastolic blood pressure, waist circumference, diabetes, diabetes medication, hypertension, hypertension medication, glucose level, triglyceride, family history of hypertension, and HDL level for the two groups (p < 0.001).

Feature selection

SBE feature selection methods were combined with ML algorithms SVM, RF, LR and XGB to identify the most significant predictors among the 30 variables considered, aiming to enhance the performance of CVD predictive models (refer to methods). Table 2 presents a comparison between the rankings of selected features based on variable importance and those used in the FRS and RPCE.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Comparison between ranked features selected by ML feature selection and the FRS risk score.

https://doi.org/10.1371/journal.pone.0323949.t002

Notably, gender, HDL, and glucose level emerged as common predictors across all ML feature selection models. Among these, gender and HDL are also integral components of the FRS and RPCE, underscoring their consistent relevance in CVD risk prediction.

Algorithm performance

On the 30% validation dataset, the ML models built using ML-selected features showed higher predictive performance compared to the FRS and RPCE risk score (Table 3). The best-performing model was the base LR model using LR-selected variables, which significantly outperformed both the FRS and RPCE risk scores (p < 0.001). A detailed performance comparison of the ML models score on a 30% validation set is provided in Table 4. For reference, the performance of the FRS and RPCE risk scores, as well as the ML models with feature selection based on a 70% training dataset, is available in S2 Table.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. The AUC of FRS and RPCE risk scores and ML models with feature selection based on a 30% validation dataset.

https://doi.org/10.1371/journal.pone.0323949.t003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Detailed performance metrics of ML models with feature selection.

https://doi.org/10.1371/journal.pone.0323949.t004

The prediction performance of the ML models varied based on the feature selection approach used. Among the individual ML models, the base LR model using features selected by LR demonstrated the highest predictive performance with an AUC of 0.769 (95% CI: 0.708–0.829). Several other base models, particularly those employing LR or SVM for feature selection, achieved comparable performance. However, base models such as SVM (radial kernel) and neural network generally exhibited lower predictive ability. The EL models, designed to combine the strengths of multiple base models, did not consistently outperform the best individual models and, in some cases, performed worse, suggesting that the ensemble approach did not necessarily improve predictive accuracy in this specific context.

When compared to traditional risk rankings, machine learning models produced promising results. The FRS and RPCE risk scores had AUC values of 0.716 and 0.740, respectively. The top-performing ML models, including the base LR with LR-selected features, performed similarly or better than the established risk scores. This shows that ML techniques, particularly when combined with proper feature selection methods, have the potential to improve predictive accuracy beyond what is currently provided by clinical scoring systems such as FRS and RPCE. However, the heterogeneity in performance between ML models and feature selection methodologies highlights the significance of thorough model selection and validation..

Secondary analysis on best performing model

The best performing ML models, base LR (LR selected var), were also trained on an imputed dataset and evaluated utilizing the 30% complete cases validation dataset. This enables a valid comparison between models constructed with imputed and complete cases models (Table 5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Detailed performance metrics of best LR model (LR selected var) on the imputed dataset and complete dataset.

https://doi.org/10.1371/journal.pone.0323949.t005

LR (LR selected var), trained on imputed datasets performed comparably to models trained on the complete dataset using a similar validation dataset of complete cases: LR (LR selected var) (AUC = 0.768, CI: 0.708–0.828 vs AUC = 0.769, CI: 0.708–0.829) (p = 0.0178). There is no statistically significant difference between the LR model (LR selected var) using complete cases with the imputed model.

Model interpretation

SHAP analysis.

The SHAP summary plots demonstrate the distribution of SHAP values for the significant features identified through LR-based feature selection. These features are used to predict CVD outcomes for Malaysian population.

The vertical axis of each plot displays the features in descending order of their influence on the model’s predictions. The x-axis corresponds to the SHAP value, which measures the influence of each feature on the model’s output. The colour gradient indicates the original value of the feature, with red representing higher values and blue representing lower values.

The SHAP analysis shown in Fig 2 identifies key variables impacting CVD occurrence, demonstrating the direction and extent of each feature’s impact. Systolic blood pressure has the most influence, with higher values (red) closely associated with positive SHAP observations, indicating an increased risk of a CVD event. Weight shows a similar positive association, with increased weight resulting in higher estimated risk. Waist circumference, another important measure of cardiovascular health, follows a similar pattern. A positive family history of hypertension, as well as a diabetes diagnosis, both increase the predicted risk. Baseline HDL cholesterol reveals an inverse connection, with higher HDL levels (red) associated with decreased risk, which is consistent with clinical experience. BMI enhances the information provided by weight and waist circumference. While gender appears to have some influence, it is less significant than the other characteristics. Smoking status has a moderate impact, with current or past smokers having a higher risk. LDL cholesterol and glucose levels have some influence, though not as much as the top-ranked features. A family history of cancer has a relatively minor impact. These variables, among others, highlight the complex nature of CVD risk, which encompasses demographic, clinical, and lifestyle factors.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. SHAP Analysis.

https://doi.org/10.1371/journal.pone.0323949.g002

NRI analysis.

NRI for the CVD model, the net reclassification of Malaysian population using the base LR (LR selected var) produced a net reclassification improvement of 13.15% and 7.00 with p < 0.00001 over the original FRS (Table 6) and RPCE (Table 7) risk score respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. NRI Comparison between Machine Learning Methods and FRS.

https://doi.org/10.1371/journal.pone.0323949.t006

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. NRI Comparison between Machine Learning Methods and RPCE.

https://doi.org/10.1371/journal.pone.0323949.t007

Discussion

The findings of this study underscore the effectiveness of ML models in predicting CVD risk, particularly when compared to traditional risk scores such as the FRS and the RPCE. Among the ML models tested, the LR model, especially with LR-selected features, demonstrated the best performance, achieving an AUC of 0.77. This result surpasses the performance of both the FRS and RPCE risk scores, highlighting the potential of ML to enhance predictive accuracy for CVD risk, especially within the diverse Malaysian population.

The superior performance of the LR model can be attributed to several key factors. LR is inherently well-suited for binary classification tasks like predicting CVD events, making it a robust baseline model. The features selected for the LR model, such as gender, systolic blood pressure, HDL cholesterol, glucose levels, and smoking status, were identified using LR and sequential backward elimination methods. These variables are well-established risk factors in the literature [54–56] and play critical roles in cardiovascular health, directly influencing outcomes like blood pressure and lipid metabolism. The inclusion of these features likely contributed to the enhanced predictive power of the LR model, as each plays a significant role in cardiovascular health.

The SHAP analysis conducted in this study provided valuable insights into the key predictors of CVD risk identified by the LR model. The SHAP summary plot illustrates the global contribution of each input feature to the model’s predictions, highlighting the importance of various demographic, clinical, and lifestyle factors in determining CVD risk.

Systolic blood pressure (SBP), weight, and waist size are all significant predictors of CVD risk. Elevated SBP has repeatedly been related to an increased risk of cardiovascular events, with studies indicating that every 10 mm Hg increase in SBP increases the risk of atherosclerotic cardiovascular disease by 53% [57]. This association is continuous and independent, with lower SBP linked with significantly lower cardiovascular mortality [58]. Similarly, weight and waist circumference are key indices of cardiovascular health; studies demonstrate that weight growth and greater waist measurements are associated with an increased risk of coronary heart disease and death. A study discovered that both body mass index (BMI) and waist circumference independently predict higher cardiovascular risk, emphasising the importance of waist circumference in women [59]. Collectively, these variables highlight the necessity of maintaining good blood pressure and body composition in order to reduce CVD risk.

A positive family history of hypertension considerably increases the expected risk of CVD. According to research, family history is an independent predictor of myocardial infarction and stroke, especially in hypertensive individuals [60]. This familial link emphasises the importance of proactive surveillance in at-risk populations. Furthermore, diabetes diagnosis is an important determinant in determining CVD risk due to accompanying metabolic abnormalities. Diabetes has regularly been shown in studies to enhance susceptibility to cardiovascular events; one study found that people with diabetes have a considerably higher risk of CVD events than those without diabetes [61]. This supports the SHAP analysis’s findings, which show that having a family history of hypertension and diabetes increases the predicted risk.

HDL cholesterol has an inverse connection with CVD risk; greater levels are linked to lower risk. A U-shaped connection has been reported, with both low and extremely high HDL levels increasing the risk of cardiovascular events, but moderate levels mitigating against them [62,63]. In contrast, LDL cholesterol is a major contributor to atherosclerosis but has less predictive ability than other indicators such as SBP and metabolic status. Elevated LDL levels have been linked to an increase in CVD mortality, emphasising the importance of efficient lipid management in at-risk individuals [64].

BMI improves our understanding of weight and waist circumference in relation to CVD risk. Although BMI is a common statistic for determining weight status, it might not sufficiently account for the impact of body composition on health outcomes. According to research, BMI corresponds with cardiovascular risk, especially when paired with other measurements such as waist circumference [65]. Gender differences in CVD risk are significant but less pronounced; although men tend to have greater incidences at younger ages, postmenopausal women face increasing risks due to hormonal changes. According to studies, female patients typically have higher cardiovascular risk profiles than males [66].

Smoking status has a major impact on CVD risk, with current or former smokers being more susceptible to cardiovascular events. Numerous studies demonstrate that smoking causes endothelial dysfunction and promotes atherosclerosis, resulting in higher morbidity from heart disease [67]. This emphasises smoking as a significant lifestyle component in CVD prediction.

Glucose levels contribute to CVD risk, however they have lower predictive ability than other variables. Elevated glucose has been associated to diabetes-related issues impacting heart health; however, the direct impact differs depending on individual metabolic profiles. Finally, while a family history of cancer has been identified as a small risk factor for CVD, genetic predispositions can affect a variety of health outcomes [67]. In conclusion, the interaction of these variables demonstrates the multidimensional character of CVD risk assessment. The SHAP study clearly demonstrates how demographic, clinical, and lifestyle factors influence individual cardiovascular health profiles.

The results of this study align with other research conducted in the Asian region, which also emphasizes the effectiveness of ML models over traditional risk scores in CVD risk prediction. For instance, a study from Korea [23] demonstrated that ML models, including neural networks, outperformed the FRS in predicting CVD risk, achieving an AUC of 0.80. Similarly, research conducted in China found that a Random Forest model achieved an AUC of 0.79, surpassing the FRS, which had an AUC of 0.76 [68]. A study from Japan reported that ML models provided better performance in predicting long-term CVD risk compared to conventional methods, with neural networks achieving an AUC of 0.82 [21]. These findings, combined with our study, suggest that ML models are more adaptable to the specific demographic and health characteristics of Asian populations, making them a more suitable choice for CVD risk prediction in these regions.

When comparing the performance of ML models to traditional risk scores in this study, the lower AUC values for the FRS and RPCE (0.74 and 0.76, respectively) suggest that these conventional methods may not fully capture the complexity of CVD risk factors in the Malaysian population. This result is consistent with the findings from other Asian studies, where traditional risk scores have similarly underperformed [23,69]. This underperformance highlights the necessity for developing and validating more population-specific models, such as the ML-based approaches used in this study.

A key strength of this study is the integration of the NRI as an evaluation metric. The NRI value of 13.15% indicates a significant improvement in correctly classifying individuals into appropriate risk categories compared to traditional risk scores, which is crucial for targeted intervention and improving clinical outcomes. The statistical significance of this NRI (p < 0.00001) underscores the practical benefits of using ML models in clinical practice, particularly in multiethnic populations like Malaysia’s, where risk stratification can be complex due to varying genetic, environmental, and lifestyle factors. Notably, the use of NRI in this context is novel, as it has not been widely reported in other studies focused on CVD risk prediction, highlighting the innovation and relevance of this study.

While EL techniques are often expected to outperform individual models, the ensemble models in this study did not exceed the performance of the base LR model. This suggests that while ensemble methods, such as those used in GBM and Neural Networks, can capture complex patterns in data, they may introduce additional complexity that does not necessarily translate into better performance. In contrast, the simpler LR model, combined with effective feature selection and SHAP analysis, provided a balance of accuracy and interpretability that is particularly valuable in clinical settings [28]. The SHAP-enhanced LR model offers a clear understanding of how individual features contribute to the overall risk prediction, making it more actionable for clinicians.

The Malaysian registry, which serves as the basis for this study, is another strength, as it represents a multiethnic population that is reflective of the broader South-East Asian region, including countries like Singapore, Indonesia, India, and China. This diversity enhances the generalizability of the findings, making the ML model applicable across various Asian populations with similar demographic and health characteristics. This broader applicability underscores the potential of the model to be used in different healthcare settings within the region.

This study also highlights the importance of continuous real-world validation, a key strength enabled by the development of an accessible web platform. This platform allows for ongoing validation of the algorithms against real-world data, ensuring that the model remains relevant and accurate as it is applied in clinical practice. Such a resource is critical for maintaining the model’s applicability and improving patient outcomes over time.

Furthermore, the deployment of the ML algorithm on a web platform accessible via www.myheartrisk.uitm.edu.my represents a significant advancement in making these predictive tools accessible to healthcare professionals. This platform also allows for continuous validation against real-world data, which will be crucial in monitoring the model’s performance over time and potentially identifying any drift due to changes in risk factors or population characteristics since the original data collection period. The integration of both the ML model and traditional risk scores like FRS into the platform enables continuous comparative assessment, ensuring the model’s reliability and relevance over time. This approach not only enhances the model’s clinical utility but also supports its adoption in diverse healthcare settings across the Western Pacific region.

A limitation of this study is the use of data acquired between 2007 and 2017. We recognise that lifestyle, healthcare practices, and the prevalence of CVD risk factors may have changed since this time. While this work provides an important baseline and proof-of-concept for the suggested ML models, the applicability of our findings to the current population should be approached with caution. Our data source is actively cleaning and preparing more recent data, which is not yet available for research use. Future research using this updated data will be necessary to test and develop our models, assess their performance in the present environment, and analyse the impact of any changes in CVD risk factors that may have occurred since the initial data.

A key area for future development is the adaptation of our models for shorter prediction horizons. While our study focused on the clinically relevant 10-year risk prediction time frame, we acknowledge that shorter-term projections (e.g., 3-year and 6-year risk) may be more insightful for older adults. Future work will include developing and evaluating models tailored to these shorter timescales.

In summary, the SHAP analysis conducted in this study not only validated the importance of key CVD risk factors but also provided a transparent framework for understanding how these factors influence predictions. This interpretability, combined with the competitive performance of the LR model, positions it as a valuable tool for CVD risk assessment, especially when compared to more complex

Conclusion

This study demonstrates the potential of ML models, including ensemble learning, to significantly improve CVD risk prediction in the Malaysian population, serving as a model for other countries in the Western Pacific region. The LR model with selected features emerged as the best performer, surpassing traditional risk scores like the FRS and RPCE in predictive accuracy. The incorporation of SHAP analysis provided valuable insights into the most influential predictors, enhancing the interpretability of the ML models.

In addition to developing these predictive models, we successfully deployed the best-performing algorithm on a publicly accessible web platform, allowing healthcare professionals and researchers to apply the model in real-time clinical settings. This platform ensures ongoing validation against real-world data, maintaining the model’s relevance and accuracy over time.

The findings underscore the importance of developing population-specific risk models that account for the unique genetic and environmental characteristics of the Western Pacific region. By demonstrating a significant improvement in the NRI over traditional models, this study highlights the potential of ML approaches to refine risk stratification and support more accurate and personalized clinical decision-making.

This study validates the effectiveness of ML and ensemble learning models in accurately predicting CVD risk within the Malaysian population, offering a model that can be extended to other countries in the Western Pacific region. The deployment of the algorithm on a web platform further enhances its utility, providing a practical tool for real-time clinical decision-making and continuous comparative assessment with traditional risk scores. The platform’s design ensures that the model remains relevant, reliable, and easily accessible, paving the way for improved patient outcomes and public health strategies across the region.

Supporting information

Acknowledgments

We thank the Nafiza Mat Nasir and UiTM for providing the REDISCOVER data. We also extend our gratitude to the researchers and participants involved in the study for their invaluable contributions.