Background

The emergence of machine learning (ML) technologies in the medical sector has revolutionized how diseases, particularly CAD, are predicted and managed. CAD has emerged as a primary contributor to the global burden of disease, claiming a significant number of lives annually [1]. In Taiwan, it ranks as the second most common cause of mortality across genders, as reported by the Health Promotion Administration of the Ministry of Health and Welfare in 2020. The most effective approach to mitigate or slow the progression of this disease involves the creation of a robust screening mechanism that can detect cardiovascular risk factors early on.

A plethora of factors including age, gender, obesity, elevated blood pressure levels, dyslipidemia, and glucose anomalies, along with smoking and alcohol consumption behaviors, have been universally recognized as contributors to the risk of developing CAD [2]. The pioneering Framingham Heart Study introduced a cardiovascular risk prediction model, known as the Framingham risk score, utilizing conventional risk indicators (e.g., age, gender, smoking status, HDL cholesterol levels, systolic blood pressure, treatment for hypertension, and diabetes presence) to predict the likelihood of coronary heart disease events, both fatal and non-fatal [3]. It has been previously suggested that the Framingham score encompasses a limited number of predictors and may overestimate CVD risk, potentially leading to overtreatment [4, 5]. Subsequently, several risk prediction models incorporating the aforementioned conventional factors have been formulated to pinpoint individuals at elevated risk for heart diseases [6,7,8,9,10,11]. While these models offer satisfactory risk predictions with C statistics ranging between 0.65 and 0.85 [12, 13], their derivation from populations of European or American descent raises concerns about their applicability to Asian demographics, potentially leading to inaccurate risk assessments [14,15,16,17].

The limitations inherent in these conventional cardiovascular risk prediction models, coupled with the potential for population-specific discrepancies, have been acknowledged [12, 18]. As a result, there has been interest in incorporating novel cardiovascular risk indicators (such as coronary artery calcium scores, carotid intima-media thickness, ankle-brachial index, and flow-mediated dilation) to improve the predictive accuracy of these algorithms [18]. Despite this, enhancements brought about by these novel markers have been marginal or not cost-effective.

In the face of these challenges, the deployment of artificial intelligence (AI) in healthcare, particularly in enhancing the precision of disease prediction, has seen a rapid increase [19,20,21]. Nonetheless, the particularities of CAD risk factors within the Taiwanese population have not been extensively studied. The application of AI-driven models in cardiovascular disease prediction promises to offer more nuanced risk assessments. This study aims to leverage an extensive set of predictive factors through AI algorithms, thereby enhancing risk stratification and making significant contributions towards the advancement of precision medicine. The goal herein is to discern the attributes associated with CAD and to formulate a risk prediction model tailored to the Taiwanese cohort.

Materials and methods

Study population, data source, and outcome variable

The study utilized data from Taiwan Biobank, a large-scale database containing health-related information from Taiwanese adults. These individuals were assessed between 2008 and 2020. A total of 132,720 subjects were initially included in the dataset (Fig. 1). Subjects with missing values (n = 549) were excluded, resulting in a final study population of 132,171 subjects. The inclusion criteria focused on subjects with complete data across several variables. The primary outcome variable of interest was the presence of self-reported CAD among the study participants. A total of 1,699 subjects in the dataset reported a history of CAD. Approval for this study was provided by the institutional review board (IRB) of Chung Shan Medical University (CS1-20009). As the data were de-identified, informed consent was waived by the institutional review board.

Fig. 1
figure 1

The pipeline describing the machine learning approach

Full size image

The following features were included as predictors in the cardiovascular risk prediction models: body mass index (BMI), smoking status, gender, alcohol consumption (drinking), total cholesterol (T_CHO), high-density lipoprotein cholesterol (HDL_C), low-density Lipoprotein cholesterol (LDL_C), triglycerides (TG), blood urea nitrogen (BUN), creatinine, alanine aminotransferase (ALT), systolic blood pressure (SBP), diastolic blood pressure (DBP), and age [3]. Blood pressure measurements were obtained during assessment using an automated sphygmomanometer in a seated position. Two readings were taken and the average measurements were used for analysis. Individuals who had smoked consistently for at least six months and were currently smoking were classified as current smokers. Conversely, those who had never smoked or had quit smoking were categorized as nonsmokers. Similarly, individuals who habitually consumed more than 50 ml of alcohol per week for over six months were considered drinkers, whereas those with no alcohol intake, or who had abstained from drinking for more than six months, were considered nondrinkers. During assessment, blood pressure measurements were obtained using an automated sphygmomanometer in a seated position. Two readings were taken and the average measurements were used for analysis. Lipid panel measures were obtained using standardized enzymatic colorimetric assays.

Propensity score matching

Propensity score matching was performed to balance potential confounders between subjects with and without CAD. A 1:4 matching ratio was applied (Fig. 1), resulting in a matched cohort of 8,495 subjects (1,699 with CAD and 6,796 without CAD) for subsequent analysis. This method facilitated the creation of a balanced dataset, enhancing the comparability between the CAD and no CAD groups and mitigating the influence of confounding variables. CAD status was determined based on self-reported questionnaires.

Machine learning algorithms and data partitioning

A variety of machine learning-based algorithms were employed to construct cardiovascular risk prediction models using the aforementioned variables. These algorithms included: Bayesian Network, Logistic Regression, Random Forest, Neural Network, and Gradient Boosting. The dataset was partitioned into training (80%) and testing (20%) sets. The training set was used to train the machine learning models and the testing set was used to evaluate the performance of the models.

Model training and evaluation

Each machine learning algorithm was trained on the training set using the selected predictors. Model performance was evaluated using metrics such as accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC), Youden’s index, and F1 score (a measure of the harmonic mean of precision and recall). The best-performing models were then evaluated on the independent testing set to assess their generalizability and predictive performance.

Statistical analyses

We utilized SAS® Viya® (version 3.5, SAS Institute Inc., Cary, NC, USA) to automate the AI models. The dataset was split into training (80% of the data) and test (20% of the data) sets before developing machine learning models. Model performance was evaluated using the AUC metric, which assesses the ROC curve. We considered the various supervised learning models described above. An AUC value close to 1 indicated a well-performing model. CAD was assigned as the dependent variable. Continuous variables were presented as mean ± standard deviation, and categorical variables were expressed as frequencies and percentages. The importance of predictors in the Gradient Boosting model was determined based on their relative influence on the model’s predictive performance.

Results

After excluding subjects with missing data, 1,699 individuals were identified with CAD through self-reported questionnaires, and propensity score matching yielded a final analysis set of 8,495 subjects (Table 1). The demographic and clinical features demonstrated significant distinctions between individuals with and without CAD. A larger proportion of those with CAD were men compared to women (66.69% vs. 33.31%, p < 0.001). Individuals with CAD were older on average compared to those without CAD (59.77 years vs. 49.58 years, p < 0.001). T_CHO, HDL_C, LDL_C, and TG were all significantly higher among individuals with CAD compared to those without CAD (p < 0.001 for all). A higher percentage of individuals with CAD were smokers and alcohol drinkers. Renal and liver function markers were also higher among individuals with CAD.

Table 1 Demographic features of the study population
Full size table

The variable importance scores for the gradient-boosting champion model are displayed in Fig. 2. Among the 14 most influential features impacting the prediction of CAD, age emerged as the most relevant variable. This underscores the importance of age as a critical factor in the gradient-boosting model’s decision-making process, highlighting its relevance in CAD risk prediction within the studied dataset.

Fig. 2
figure 2

This plot shows the 14 most important variables, as determined by the Gradient Boosting (champion) model. The most important input for this model was age, followed by T_CHO.

Full size image

Table 2 summarizes the performance metrics of various machine learning models in predicting CAD risk. The evaluation of predictive models indicated varied performances across different metrics. The Gradient Boosting model showcased the highest AUC value of 0.846, with a 95% CI of 0.819 to 0.873, suggesting it was the most effective in distinguishing between the classes. However, both the Bayesian Network and Random Forest models achieved the highest sensitivity, at 0.794 (95% CI: 0.751–0.837), indicating their precision in identifying true positives. Specificity was led by the Gradient Boosting model, reaching 0.759 (95% CI: 0.736–0.782), which denotes its strength in correctly identifying true negatives. This model also scored the highest in accuracy, with a value of 0.762 (95% CI: 0.742–0.782), and in F1 score, at 0.567 (95% CI: 0.543–0.591), reflecting its overall balanced performance in precision and recall. Logistic Regression and Neural Network models presented competitive performances with AUC values of 0.838 (95% CI: 0.811–0.865) and 0.836 (95% CI: 0.808–0.864), respectively. Although these models showed slightly lower sensitivity and specificity than the leading models, they remained robust in their predictive capabilities. The AUC-ROC curves for all models are shown in Fig. 3.

Fig. 3
figure 3

The AUROC for all models. (Gradient Boosting was the champion model)

Full size image
Table 2 Performance of predictive models under consideration
Full size table

Discussion

Principal findings

The results of our study provide valuable insights into the demographic characteristics, risk factors, and predictive performance of machine learning models in assessing CAD risk in the studied population. Our analysis encompassed a range of performance metrics to evaluate the efficacy of different machine-learning algorithms. The gradient-boosting champion model emerged as the most effective in predicting CAD risk, achieving an AUC of 0.846. This high AUC value indicates the model’s strong discriminatory power in distinguishing between CAD-positive and CAD-negative cases. This is particularly notable as the value falls within the 0.8 to 0.9 range, considered accurate for predicting cardiovascular diseases with machine learning [5] In contrast, results from a previous study assessing atherosclerotic cardiovascular disease in Taiwan [22] showed that the eXtreme Gradient Boosting (XGBoost) and random forest models demonstrated the best performance with AUC-ROC values of 0.72 (0.68–0.76) and 0.73 (0.69–0.77) respectively, though not significantly better than other models.

Our study also showed solid results across other metrics, such as sensitivity, specificity, accuracy, and F1 Score, showcasing its reliability in CAD risk prediction, with AUC values ranging from 0.825 to 0.838. These models demonstrated varying degrees of sensitivity, specificity, accuracy, and F1 Score, indicating their differential capabilities in capturing CAD-related patterns and making accurate predictions. Based on prior research findings [23], future improvements in predicting recurrent cardiovascular disease risk may come from using comprehensive datasets and employing advanced, interpretable AI models, which could enhance precision and maintain clarity in decision-making processes. The adoption of AI models presents an opportunity to augment risk prediction capabilities. These strategic approaches signify potential pathways for advancing the precision and efficacy of cardiovascular event risk prediction in future research endeavors.

Our results further reveal that people with CAD often had higher risk factors such as age, BMI, high blood pressure, and poor lipid and renal function. Age was identified as a significant predictor of the disease. Our analysis also uncovered demographic differences in CAD prevalence, with men at higher risk, and lifestyle factors like smoking and drinking significantly affecting CAD risk. This underlines the need for lifestyle changes in CAD prevention strategies.

In Taiwan, the application of ML models for predicting CVD risk is gaining attention due to its potential to tailor preventive strategies and improve patient outcomes [24]. The country’s unique healthcare infrastructure, characterized by its National Health Insurance (NHI) system and TWB, offers extensive patient data, making it an ideal environment for testing these advanced predictive tools. In our study, we employed a neural network model with specific parameters designed to optimize performance while maintaining simplicity and interpretability. The architecture of the neural network consisted of a single hidden layer comprising 50 neurons. We utilized the hyperbolic tangent (Tanh) function as the activation function for the hidden layer due to its ability to introduce non-linearity and its effectiveness in handling a wide range of input values. The optimization of the network’s weights was performed using the Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm, chosen for its efficiency in handling large-scale optimization problems and its suitability for neural network training. By explicitly detailing the neural network parameters, we aim to provide a clear framework that can be readily reproduced and built upon by future researchers. This transparency not only enhances the reproducibility of our findings but also facilitates a deeper understanding of the model’s behavior and performance characteristics.

Traditional algorithms for predicting cardiovascular disease have shown varying degrees of accuracy, with c statistics ranging from 0.65 to 0.85 [12, 13]. However, the integration of machine learning (ML) into healthcare for predicting CAD risk is showing promising results, with a notable increase in popularity due to its potential for more accurate predictions. A significant study utilizing data from the Multi-Ethnic Study of Atherosclerosis highlighted that ML algorithms outperformed both Cox proportional hazard models and traditional risk scores in CAD risk prediction [25]. Further research [23, 26] supports the advantage of ML in enhancing the accuracy of cardiovascular risk models through improved discrimination and calibration.

Previous explorations into CAD risk prediction have also ventured into the realm of genetic markers. One study introduced a combination of traditional risk factors, novel biomarkers, and a comprehensive set of genetic markers into ML models to predict coronary artery calcification [27]. Despite these efforts, the results yielded sensitivity and specificity rates of approximately 70% and 60%, respectively, suggesting that the addition of genetic data may not inherently boost prediction accuracy. The evidence suggests that ML algorithms may effectively harness traditional risk factors for CAD in the presence or absence of absence of new markers [23].

While much of the existing literature on machine learning in cardiovascular disease has focused on imaging-based approaches, routinely collected clinical biochemical indicators represent an important and underexplored area. A recent study has demonstrated the potential of machine learning models utilizing clinical data, such as blood biomarkers, to predict the presence and risk of cardiovascular diseases [28]. The authors developed a machine learning model based on 13 features, including lipid panel measures, to accurately identify individuals with coronary artery disease. Our findings add to this emerging body of research, highlighting the value of leveraging readily available clinical data for machine learning-based cardiovascular risk assessment. By constructing predictive models using common biochemical indicators, we can potentially provide a cost-effective and scalable approach to supporting clinical decision-making, complementing or even outperforming more resource-intensive imaging-based techniques in certain settings.

Strengths and limitations

While our study points to the potential of machine learning in enhancing CAD risk prediction, we acknowledge its limitations, including its retrospective design and the need for further validation [21]. Furthermore, our investigation was hindered by a deficiency in data concerning disease severity within our study questionnaires. Consequently, we were unable to ascertain this crucial aspect. Finally, the CAD diagnosis was determined solely based on participants’ responses indicating they had ever been diagnosed with CAD by a doctor. We could not cross-reference this self-reported data with medical records or claims data from other data sources. The lack of objective clinical confirmation of the disease diagnosis might have introduced the potential for inaccuracies or biases. Participants may have under-reported or over-reported their CAD history, which could impact the reliability of our findings. Future research should aim to validate self-reported disease status against medical documentation to strengthen confidence in the results. Despite these shortcomings, our research underscores the value of machine learning, especially gradient boosting models, in providing accurate CAD risk assessments, which could improve clinical practices for early intervention and personalized care.

Conclusions

In conclusion, these findings suggest that the Gradient Boosting model performed well in discriminating between CAD-positive and CAD-negative cases within a Taiwanese cohort, making it a promising tool for CAD risk prediction. Identifying key predictors supports the potential of targeted interventions and personalized medicine approaches in managing and preventing CAD.