Interpretable predictive value of including HDL-2b and HDL-3 in an explainable boosting machine model for multiclass classification of coronary artery stenosis severity in acute myocardial infarction patients

Abstract

Aims

The aim of this study was to use explainable boosting machine (EBM) to evaluate the predictive value of HDL-2b and HDL-3 levels in comparison with traditional lipid parameters in three-class classification of coronary artery stenosis severity in acute myocardial infarction (AMI) patients.

Methods and results

In this cross-sectional study, 1200 AMI patients were evaluated. HDL subtypes were quantified via microfluidic chip detection, and stenosis severity was assessed via the Gensini scoring system. The Gensini scores were divided into three groups: low group (<36.5), moderate group (36.5–72), and high group (>72). Explainable boosting machine, an interpretable machine learning technique, was employed to assess the predictive value of HDL-2b and HDL-3 compared with traditional lipid markers. Explainable boosting machine was used as the main model in this study, whereas logistic regression, XGBoost, and Random Forest were selected as reference models for predictive performance. Model performance was evaluated using receiver operating characteristic curves. The HDL-3 (%) values were divided into three risk categories: low (>43), moderate (30–43), and high (<30). The incorporation of HDL-2b and HDL-3 levels into lipid profiling significantly increased the group importance scores. The macro-average area under the curve values for the four models were as follows: 0.56 for the logistic model, 0.54 for the EBM model, 0.50 for the Random Forest model, and 0.49 for the XGBoost model.

Conclusion

HDL-3 provides superior predictive value for evaluating coronary artery stenosis severity in AMI patients compared to HDL-2b and other conventional lipid markers.

Graphical Abstract

Graphical Abstract.

Introduction

Acute myocardial infarction (AMI) remains a leading cause of morbidity and mortality worldwide, necessitating improved diagnostic and predictive tools to manage this condition effectively. Traditional lipid parameters, such as triglyceride (TG), total cholesterol (TC), high density lipoprotein cholesterol (HDL-C), and low density lipoprotein cholesterol (LDL-C), are commonly used in clinical practice to evaluate cardiovascular risk. However, these conventional markers often provide limited insight into the complexity of lipid metabolism and its impact on coronary artery disease (CAD) progression.

HDL is known for its cardioprotective properties, primarily due to its role in reverse cholesterol transport. HDL particles are heterogeneous, comprising several subtypes with distinct physiological functions.^1,2 Among these, HDL-2b and HDL-3 have garnered significant attention. HDL-2b is typically larger and more buoyant and is associated with enhanced reverse cholesterol transport, whereas HDL-3 is smaller and denser and is often linked with antioxidative and anti-inflammatory properties.^1,2

Recent studies have indicated that the distribution of HDL subtypes varies significantly among individuals with different cardiovascular conditions.³ For example, patients with acute coronary syndrome and stable CAD exhibit distinct HDL-2b and HDL-3 profiles, suggesting that these subtypes may play a role in disease progression and severity assessment.³ Despite these findings, the clinical utility of HDL-2b and HDL-3 levels in predicting the severity of coronary artery stenosis in AMI patients remains underexplored.

The aim of this study was to use explainable boosting machine (EBM),⁴ an interpretable machine learning model, to evaluate the predictive value of HDL-2b and HDL-3 levels in comparison with traditional lipid parameters in three-class classification of coronary artery stenosis severity in AMI patients.

Methods

Study participants

This study was a single-centre, cross-sectional analysis. Adult patients initially admitted for AMI (ICD-9 code: 410) at the cardiovascular centre of Beijing Tsinghua Changgung Hospital, from September 2017 to June 2023, were retrospectively identified from the hospital’s electronic health record (EHR) system. Patients who did not undergo coronary angiography were excluded. This study was conducted in compliance with the Helsinki Declaration and was approved by the ethics committee of Beijing Tsinghua Changgung Hospital (number 23578-4-01). In consideration of the retrospective nature of the study, the requirement for patient informed consent was waived.

Definition of variables

Demographic information, medical history, laboratory test results (including traditional lipid profiles, routine blood tests, and myocardial injury markers, etc.), echocardiography reports, and coronary angiography procedural records were extracted from the EHR system. Blood sample residuals from laboratory testing during the patient’s hospitalization were preserved at −80°C in a biobank at Beijing Tsinghua Changgung Hospital. All HDL subtypes were evaluated at Guangdong Ardent BioMed Co., Ltd (ABIOM), Guangzhou, China, utilizing ABIOM’s HDL Subfraction Assay Kit and the Microfluidic System (MICEP-30) in accordance with the manufacturer’s guidelines.⁵ Blood samples underwent centrifugation to separate serum, which was subsequently stored at −20°C and analysed within 7 days to maintain sample integrity. Serum samples were incubated for 5 min with a sample solution containing a proprietary lipid-specific dye that specifically stains HDL particles. HDL subtyping was performed using chip capillary electrophoresis, a technique that separates particles according to their molecular size and charge characteristics. The sample was introduced into the separation channel of the microfluidic chip, preloaded with an electrophoretic medium, via electroosmotic flow under an applied electric field. Fluorescence intensity profiles, representative of HDL subtypes, were documented in relation to migration time. The electrophoretic patterns of HDL subtypes were further examined by deconvolution.⁶ The units of measurement for HDL-2b and HDL-3 are percentages (%). Due to the limitations of the test kit, only HDL-2b and HDL-3 were quantifiable in this investigation.

The four classic lipid profiles include TG, TC, HDL-C, and LDL-C concentrations. The severity of coronary stenosis was determined via the Gensini scoring system. The dependent variable was the Gensini group, a multiclass classification target divided into tertiles based on the Gensini score. The three Gensini score categories for the classification of stenosis severity were: low group (<36.5), moderate group (36.5–72), and high group (>72).

Explainable boosting machine

Explainable boosting machine represents a notable development in machine learning,⁴ and it strikes a unique balance between interpretability and predictive accuracy. Explainable boosting machine uses a tree-based cyclic gradient boosting technique, a type of generalized additive model, to automatically detect interactions.⁷ Moreover, EBM provides transparent insights into the underlying decision-making processes, which are critical in clinical settings where understanding predictions is essential, in contrast to traditional black-box models. To capture nonlinear relationships and model interactions when performance is significantly improved, EBM iteratively fits each feature with a shape function to minimize error. The application of EBM has been demonstrated in numerous clinical domains where interpretability is of paramount importance. For example, EBM was used to assess the utility of predicting angina pectoris⁸ and stroke⁹ in comparison with traditional machine learning models. One study that assessed student performance in an educational context illustrated the application of EBM to a multiclass classification task.¹⁰ A detailed explanation of the formulas used in the EBM algorithm in the context of multiclass classification can be found in previous literature.¹⁰

Model construction and evaluation

The dataset was partitioned into two subsets: a training set comprising 70% of the data and a testing set comprising the remaining 30%. Echocardiographic indicators and additional laboratory markers were excluded from this study to preserve the predictive validity of comparing HDL subtypes and traditional lipid variables regarding the severity of coronary artery stenosis. Only fundamental demographic data and comorbidities were incorporated. The independent variables included age; sex; hypertension, diabetes, stroke, kidney disease, thyroid dysfunction, and chronic obstructive pulmonary disease status; and TG, TC, HDL-C, LDL-C, HDL-2b, and HDL-3 levels. The three-class Gensini group was used as the dependent variable. It was determined that feature scaling was unnecessary. Explainable boosting machine was used as the main model in this study, whereas logistic regression, XGBoost, and Random Forest were selected as reference models for predictive performance. A grid search combined with 10-fold cross-validation was used to optimize the hyperparameters for each model. The hyperparameters were selected based on a synthesis of computational efficiency and enhancement of model performance. Furthermore, HDL-2b and HDL-3 were removed from the preliminary model to assess their influence on the predictive performance of the model.

The regularization strength parameter ‘C’ for logistic regression was adjusted on a logarithmic scale with values of [0.001, 0.01, 0.1, 1, 10, 100, 1000]. This range was selected to assess the trade-off between bias and variance, where lower ‘C’ values impose greater regularization to mitigate overfitting, but higher ‘C’ values diminish regularization to permit more adaptable fitting.

The maximum tree depth parameter (‘max_depth’) for XGBoost was examined using the values [3, 4, 5]. The selected values regulate the complexity of individual trees, with a reduced ‘max_depth’ limiting tree depth values mitigate overfitting risk, while bigger values facilitate the capture of more complex patterns in the data.

The ‘max_depth’ hyperparameter for the Random Forest model was configured at ([80, 90, 100, 110]) to evaluate the impact of tree depth on model generalizability and robustness. By limiting tree depth, we reduced the hazards of overfitting in situations with increased tree counts, while deeper configurations facilitated the investigation of potentially intricate linkages within the feature space.

The ‘greedy_ratio’ parameter for the EBM model was adjusted throughout a comprehensive range of values: ([0.0, 0.25, 0.5, 0.75, 1.0, 1.25, 1.5, 1.75, 2.0, 4.0]). The ‘greedy_ratio’ adjusted the level of aggressiveness in selecting feature interactions, influencing both the learning rate and the complexity of feature relationships. Reduced values promote cautious learning to mitigate overfitting and elevated values facilitate more assertive fitting to identify intricate patterns.

The assessment of the model’s performance involved the use of receiver operating characteristic (ROC) curves and the corresponding area under the curve (AUC). As ROC curves are commonly used in binary classification, this study used binarization of the output (per class) and one-vs.-rest (micro score) or one-vs.-all (macro score) strategies for multiclass classification. The micro-average ROC curve was plotted as the sum of all true positives and false positives across all classes. The macro-average ROC curve, which simply takes the average of curves across all classes, was also drawn.

Global and local interpretation of the classification explainable boosting machine model

Explainable boosting machine can be used to generate global explanations via the training dataset. To apply the classification EBM model, which is similar to logistic regression, the logistic link function must be used. To learn logit shape functions per feature for each class in the multiclass classification configuration, EBM employs machine learning techniques such as bagging and gradient boosting. This function allows models to be trained in logit space, which ensures that the effect of each feature is additive. The results are then converted back to bounded probabilities during prediction. The importance of a feature is determined by calculating the mean absolute value of the predicted score for each predictor. A higher score indicates a higher level of importance. In the local interpretation of test data predictions, the score assigned to each predictor variable represents the positive or negative contribution that contributes to the outcome.

Importance of feature groups in the classification explainable boosting machine model

The group importance calculations determine the significance of the feature groups in the model. Notably, overlapping features may belong to multiple groups. This study used permutations of six lipid parameters (HDL-2b, HDL-3, TC, TG, HDL-C, and LDL-C) to create new groups for importance assessment. Groups of interest were evaluated alongside other groups, and the original features were combined to determine variable significance.

Statistical analysis

All analyses were conducted via Python (version 3.11.5). Descriptive statistics were used to characterize the data.¹¹ The ‘InterpretML’ package was used to create the classification EBM model.⁴ The ‘InterpretML’ library currently supports variable interactions only for regression and binary classification tasks and has not yet extended this functionality to multiclass classification tasks. The ‘scikit-learn’ library was used for the execution of machine learning processes. The ‘Yellowbrick’ library was utilized for the visualization of performance evaluations of machine learning models.¹² Logistic regression, by definition, can only classify two discrete groups. However, in the scikit-learn library, the ‘LogisticRegression’ class can be modified to analyse data for multiple classes by changing the ‘multi_class’ argument to ‘multinomial.’ Furthermore, the ‘solver’ argument must specify the solver used, which should be compatible with multinomial logistic regression, such as ‘lbfgs.’ Data deficiencies in this study originated from indicators of the traditional lipid variables (TC, TG, LDL-C, HDL-C) because the patients did not have lipid laboratory tests during their hospitalization. Given the few instances of missing data, participants with missing values were eliminated from the analysis instead of employing data imputation.

Results

Characteristics of the participants

Supplementary material online, Figure S1 shows a flowchart of the participant screening procedure and Supplementary material online, Table S1 presents the characteristics of all patients who had coronary angiography, including instances with incomplete data. The study included 1200 participants in the formal machine learning analysis. The study population was divided into three groups according to the coronary artery stenosis Gensini score: low, moderate, and high. The baseline characteristics of this group are shown in Table 1. The median age of all study participants was 63 years (Q1, Q3: 55–72 years). This age distribution was consistent across all groups. The sex distribution revealed a male predominance in all the groups. Overall, 76.0% of the patients were male. HDL-2b levels, represented as a percentage (%), had a median value of 22.7% (Q1, Q3: 19.0–26.7%) in the overall sample. The distribution of HDL-2b slightly varied among the groups. HDL-3 levels also varied modestly across the different severity groups.

Table 1.

Characteristics of the study population with complete data (participants included in data analysis, n = 1200)

Variables	Overall (n = 1200)	Low group (n = 404)	Moderate group (n = 400)	High group (n = 396)
Age, years	63.0 [55.0,72.0]	64.0 [55.8,71.0]	62.0 [54.0,71.0]	64.0 [56.8,73.0]
Sex, female, n (%)	288 (24.0)	103 (25.5)	92 (23.0)	93 (23.5)
Hypertension, n (%)	829 (69.1)	287 (71.0)	272 (68.0)	270 (68.2)
Diabetes, n (%)	500 (41.7)	162 (40.1)	155 (38.8)	183 (46.2)
Stroke, n (%)	166 (13.8)	59 (14.6)	46 (11.5)	61 (15.4)
Kidney disease, n (%)	190 (15.8)	59 (14.6)	60 (15.0)	71 (17.9)
Thyroid dysfunction, n (%)	70 (5.8)	30 (7.4)	24 (6.0)	16 (4.0)
Chronic obstructive pulmonary disease, n (%)	34 (2.8)	15 (3.7)	9 (2.2)	10 (2.5)
Total cholesterol, mmol/L	4.3 [3.6,5.1]	4.3 [3.5,5.0]	4.4 [3.7,5.1]	4.3 [3.5,5.1]
Triglyceride, mmol/L	1.5 [1.1,2.1]	1.5 [1.1,2.2]	1.5 [1.1,2.2]	1.4 [1.1,2.0]
Low density lipoprotein cholesterol, mmol/L	2.7 [2.1,3.5]	2.7 [2.0,3.4]	2.9 [2.2,3.5]	2.7 [2.2,3.6]
HDL cholesterol, mmol/L	0.9 [0.8,1.1]	0.9 [0.8,1.1]	0.9 [0.8,1.1]	0.9 [0.8,1.1]
HDL-2b, %	22.7 [19.0,26.7]	22.6 [19.2,27.3]	22.9 [19.6,26.6]	22.7 [18.3,26.7]
HDL-3, %	35.0 [28.0,41.0]	36.0 [30.0,41.0]	36.0 [29.0,41.0]	34.0 [26.0,40.0]
Total Gensini score	52.0 [28.0,83.6]	21.0 [10.9,28.0]	53.0 [44.0,61.0]	95.0 [84.0,120.0]

The values are shown as median [Q1, Q3] or n (%). The units of measurement for HDL-2b and HDL-3 are expressed as percentages (%).

HDL, high density lipoprotein.

Global feature importance of the classification explainable boosting machine model

The global feature importance for Gensini group prediction via a classification EBM model is shown in Figure 1. The model’s relative importance for each feature is indicated by the graph, which displays the mean absolute score (weighted) for each feature. With an importance score of 0.021, diabetes status was the most significant characteristic, followed by HDL-3 (0.019). The importance scores for LDL-C (0.013), TC (0.011), TG (0.010), and HDL-2b (0.009) levels were very similar.

Figure 1.

Global feature importance of the classification EBM model. COPD, chronic obstructive pulmonary disease; HDL, high density lipoprotein; HDL-C, high density lipoprotein cholesterol; LDL-C, low density lipoprotein cholesterol; TC, total cholesterol; TG, triglyceride.

Global explanations

Based on the trend lines predicting the three-class Gensini group (Figure 2), the HDL-2b (%) values were divided into three risk categories: low risk (>28), moderate risk (17–28), and high risk (<17). When the HDL-2b (%) value was <17, the high group’s prediction curves for the EBM model were consistently above those of the other two categories. The moderate group had a significantly greater prediction curve than the other two categories did when the HDL-2b (%) values were between 21 and 28. High levels of flattening and intertwining were observed in the prediction curves for the three categories when the HDL-2b (%) value exceeded 28. As illustrated in Figure 2, the HDL-3 values (%) < 30 were found to differentiate the prediction curves of the low and high groups to a greater extent. The HDL-3 values (%) exceeding 43 were found to differentiate the prediction curves of the low and moderate. Therefore, the HDL-3 (%) values were divided into three risk categories: low (>43), moderate (30–43), and high (<30).

Figure 2.

Global explanation of HDL-2b and HDL-3 values for the prediction of the multiclass Gensini group. A positive value of the importance score indicates a positive contribution to the prediction of the three-class Gensini group, whereas a negative value indicates a negative contribution. ‘Density’ is a histogram showing the distribution of the data. HDL, high density lipoprotein.

When the concentration of LDL-C exceeded 1.5 mmol/L, the curve for the low group started to decline, while the curve for the moderate group began to ascend from a flat baseline (Figure 3). The initial intersection of the low and high group curves was observed at an LDL-C concentration of 2.1 mmol/L. Similarly, the curves for the moderate and high groups intersected first at an LDL-C concentration of 2.5 mmol/L. When the LDL-C concentration exceeded 3.0 mmol/L, the prediction curve for the high group was consistently greater than that for the low group. However, when the LDL-C concentration exceeded 5 mmol/L, the model became unstable, necessitating caution in interpreting the results due to the extremely sparse distribution of data points. Based on the trend lines of the low and high groups, the LDL-C values (mmol/L) were divided into three risk categories: low (<2.1), moderate (2.1–3.0), and high (>3.0). The predicted curves for TC concentrations (Figure 3) exhibited a comparable pattern to those observed for LDL-C values.

Figure 3.

Global explanation of low density lipoprotein cholesterol (LDL-C) and total cholesterol (TC) concentrations for the prediction of the multiclass Gensini group. LDL-C, low density lipoprotein cholesterol; TC, total cholesterol.

The importance of groups of features comprising varying numbers of elements is illustrated in Supplementary material online, Figure S2. As more group variables are included, it is evident that the combination of HDL-3 and HDL-2b values with other lipid metrics results in increased group importance scores.

Local explanations

An example of a patient with an incorrect prediction was used to understand the rationale behind the model’s predictions. An instance of a three-class Gensini group prediction for a specific patient is provided in Figure 4 and Table 2. The specific contribution of each feature to the prediction of the three categories is shown in Figure 4. The underlying data for the visualization results shown in Figure 4 are presented in Table 2. The total score is the sum of the contribution scores for each variable and intercept. The predicted probability is the result of the model transforming the score via the Softmax function, which yields the probability of the given input. For multiclass classification, the model chooses the class with the highest probability as its output. As shown in Table 2, the low category has the highest probability, indicating that the model predicted the patient’s category to be low group. However, as shown in Figure 4, the patient’s actual category was high.

Figure 4.

Local explanation of a multiclass Gensini group prediction outcome. For this patient, the EBM model predicts a probability of 0.343 for the low category and a probability of 0.317 for the high category. The model predicts low category, whereas the patient’s actual category is high. COPD, chronic obstructive pulmonary disease; HDL-C, high density lipoprotein cholesterol; LDL-C, low density lipoprotein cholesterol; TC, total cholesterol; TG, triglyceride.

Table 2.

Examples of feature importance scores for different categories of selected patients

Items	Values	High group	Low group	Moderate group
Age (years)	67	0.011	−0.005	−0.006
Sex	Female	0.005	0.023	−0.028
Hypertension status	No	0.004	−0.010	0.006
Diabetes status	No	−0.028	0.014	0.013
Stroke status	No	−0.006	0.003	0.003
Kidney disease status	No	−0.007	0.002	0.006
Thyroid dysfunction	No	0.005	−0.004	0.000
Chronic obstructive pulmonary disease status	No	0.000	−0.003	0.003
Total cholesterol, mmol/L	5.4	−0.008	−0.009	0.017
Triglyceride, mmol/L	1.64	0.006	−0.007	0.001
Low density lipoprotein cholesterol, mmol/L	3.8	0.003	−0.019	0.016
HDL cholesterol, mmol/L	1.15	0.005	0.008	−0.013
HDL-2b, %	19.62	−0.009	0.011	−0.002
HDL-3, %	35	−0.015	0.012	0.003
Intercept	1	−0.015	0.012	−0.001
Total importance score		−0.049	0.028	0.018
Predictive probability		0.317	0.343	0.340

HDL, high density lipoprotein.

Model comparisons

According to the ROC curves presented in Figure 5, the logistic model outperforms the EMB model slightly in terms of the macro-average AUC ranking. The macro-average AUC values for the four models were as follows: 0.56 for the logistic model, 0.54 for the EBM model, 0.50 for the Random Forest model, and 0.49 for the XGBoost model. A comparison of AUC metrics reveals that the difference in predictive performance among the four models is marginal.

Figure 5.

Receiver operating characteristic (ROC) curves for the different models in the test set. AUC, area under the curve; ROC, receiver operating characteristic.

Effect of HDL-2b and HDL-3 values on model performance

After the removal of the HDL-2b and HDL-3 values, a reduction in the AUC metrics of the EBM model can be observed when comparing the results presented in Figure 5 to those presented in Figure 6. Among these, the macro-average AUC decreased from 0.54–0.53. The incorporation of HDL-2b and WHDL-3 values can increase the EBM model’s AUC for low and moderate classifications, while it decreases the AUC for the high classification. Similarly, incorporating these values into the logistic model enhances the AUC across all three classifications.

Figure 6.

Receiver operating characteristic (ROC) curves for the different models in the test set, with HDL-2b and HDL-3 variables excluded from the model training and evaluation. AUC, area under the curve; ROC, receiver operating characteristic.

Discussion

Main findings

In this study, the predictive value of HDL-2b and HDL-3 values, two distinct subtypes of HDL, in assessing coronary artery stenosis in patients with AMI was demonstrated via an EBM model. When predicting the three-class Gensini group, the global importance of HDL-2b and HDL-3 values surpassed that of HDL-C concentrations. Furthermore, the assessment of group importance revealed that the variable groups containing HDL-2b and HDL-3 values were more significant than those based solely on traditional lipid metrics (TC, TG, HDL-C, and LDL-C concentrations). HDL-3 values were more effective in distinguishing between the low and high groups, whereas HDL-2b values were less effective in identifying the moderate group. Additionally, as shown in Figure 5, the AUC of 0.52 for the moderate group was among the lowest observed in the comparative assessment of the EBM model’s predictive performance across the three categories.

Clinical implications

The clinical implications of this study are substantial, suggesting that incorporating HDL subtype analysis into routine lipid profiling could enhance cardiovascular risk assessment. By utilizing the interpretable and transparent EBM model, clinicians can obtain deeper insights into individual risk factors and their contributions, potentially leading to more targeted and effective interventions for patients with AMI. Furthermore, understanding the distinct roles of HDL-2b and HDL-3 may inform therapeutic strategies aimed at modifying HDL composition to reduce cardiovascular risk. The study findings demonstrated that the cut-off points for risk stratification, which were based on the trend of the LDL-C concentration prediction curve generated by the EBM model, aligned closely with the LDL-C control targets recommended by the European Society of Cardiology (ESC) guidelines. The ESC guidelines recommend an LDL-C target of <1.4 mmol/L (<55 mg/dL) for patients at very high risk, <1.8 mmol/L (<70 mg/dL) for high-risk patients, <2.6 mmol/L (<100 mg/dL) for patients at moderate risk, and <3.0 mmol/L (<116 mg/dL) for low-risk patients.¹³ The LDL-C concentrations measured in this study had several significant cut-off points: 1.5 mmol/L, 2.1 mmol/L, 2.5 mmol/L, and 3.0 mmol/L, as detailed in the results section. The cut-off points for LDL-C risk stratification derived from the EBM model’s prediction curves are remarkably close to the values recommended by the ESC guidelines. This successful prediction underscores the unique advantages of the EBM model for evaluating the predictive value of clinical markers. Moreover, the risk stratification results obtained for HDL-2b and HDL-3 via this method are credible.

Sonora Quest Laboratories, a division of Laboratory Sciences of Arizona, operates as a joint venture between Banner Health and Quest Diagnostics, creating one of the largest integrated laboratory systems in the USA. Sonora Quest Laboratories, accredited by the College of American Pathologists, ensures the accuracy of test results, which is crucial for precise patient diagnosis. Sonora Quest Laboratories categorizes HDL-2b readings for women into three risk classifications: low risk (>28%), moderate risk (18–28%), and high risk (<18%).¹⁴ Men’s risk categories are delineated as low risk (>26%), moderate risk (18–26%), and high risk (<18%).¹⁴ This study revealed that the HDL-2b risk stratification, formulated using the EBM model based on data from the Chinese population, closely aligned with the reference intervals given by Sonora Quest Laboratories.

Machine learning in lipid risk stratification

Machine learning has demonstrated significant efficacy in discovering the most predictive biomarkers. By employing the ranking of predictor variable relevance produced by these models, researchers can identify potentially significant factors and develop robust risk models.^15,16 In lipid risk stratification, researchers commonly use segmentation methods to categorize data into groups based on median, tertiles, or quartiles, etc.^17,18 Thereafter, conventional logistic regression or cox regression is used to compute effect measures, such as odds ratios or hazard ratios, relevant to the different strata.

Restricted cubic splines (RCSs) are an innovative method for examining the relationship between continuous variables and outcomes, particularly in identifying critical cut-off values for risk categorization.^19,20 Restricted cubic spline partitions continuous variables into adaptable, continuous intervals through the use of several piecewise cubic polynomials. Restricted cubic spline methods are a commonly utilized method for analysing lipid levels for risk stratification of cardiovascular disease.^17,18,21

Although RCS enhances unambiguous visualization and threshold identification, its interpretative capability may be constrained in multiclass classification scenarios or contexts necessitating higher-order interactions, as it predominantly focuses on pairwise correlations. Restricted cubic spline is particularly ineffective in capturing high-dimensional interactions, primarily concentrating on single-variable effects. Conversely, EBMs excel at modelling intricate interactions among several predictors, therefore offering a more precise representation of the underlying data dynamics. Moreover, RCS has difficulties related to flexibility, as its efficacy is affected by the quantity and positioning of knots.¹⁹ This reliance may lead to either over or undersmoothing of essential data points. Explainable boosting machines, conversely, exhibit greater adaptability, optimizing form functions for each predictor and efficiently capturing subtle trends. Moreover, RCS is susceptible to overfitting, especially with limited sample sizes, while EBM utilizes iterative fitting and cross-validation techniques that improve resilience to overfitting. Ultimately, while RCS generates interpretable curves, the determination of clinically significant cut-off points may be fairly subjective. Conversely, EBM enhances interpretability by offering comprehensive insights on feature significance and the roles of specific variables, rendering it a more efficacious instrument for clinical applications.

Role of HDL subtypes in cardiovascular disease

Greater HDL-2b values have been consistently linked with reduced CVD risk, as HDL-2b is believed to be more effective in promoting cholesterol efflux from cells, thereby reducing atherosclerotic plaque formation. Approximately 70% of studies on HDL-2b show a significant correlation between low levels of these subtypes and increased CVD risk, highlighting its potential role as a biomarker for cardiovascular health.² Nevertheless, previous studies² have indicated that HDL-3 values appear to be less clearly associated with the risk of CVD than HDL-2b values are. Moreover, in patients undergoing percutaneous coronary intervention,²² the severity of coronary stenosis was found to be predictive of HDL-2b and HDL-3b levels. These findings underscore the potential clinical importance of assessing HDL subtypes in cardiovascular risk stratification and management.

The importance of HDL subtypes identified by microfluidic chip technology in cardiovascular health has been emphasized by recent studies. A higher risk of coronary artery stenosis has been linked to lower levels of HDL-2b and HDL-3 in patients with AMI.²³ HDL-2b offers significant diagnostic value for identifying insulin resistance and metabolic syndrome compared with conventional lipid markers.²⁴ Similarly, HDL-2b levels may serve as a predictive marker for detecting vulnerable intracranial atherosclerotic plaques.⁶

The impact of plasma TG levels and apolipoproteins on HDL subtype distribution further highlights the multifactorial nature of lipid metabolism. Studies have demonstrated that alterations in the apoB-100/A-I ratio can influence the distribution of HDL subtypes, with implications for cardiovascular risk assessment.²⁵ Furthermore, the activity of lecithin–cholesterol acyltransferase and its correlation with HDL-cholesterol and subtype levels underscore the intricate interplay between enzymes and lipoproteins in lipid metabolism.²⁶

Limitations

This study has several limitations. First, the cross-sectional design does not allow for inference of causality between HDL subtypes and the severity of coronary artery stenosis. Longitudinal studies are necessary to establish temporal relationships and causal effects. Second, this study was conducted at a single centre, which may limit the generalizability of the findings to broader populations. The four models in this investigation exhibited comparatively low AUC, each remaining below 0.7. Although this may elicit apprehensions about the models’ overall performance, it is essential to emphasize that the principal aim of this research was to acquire risk classification of HDL subtypes rather than for maximizing predictive AUC. The HDL-2b hazard stratification results from this study closely align with the HDL-2b reference ranges provided by Sonora Quest Laboratories in the USA, indicating the robustness and generalizability of the EBM model. Identifying risk classification for HDL-2b and HDL-3 offers a more thorough method for assisting physicians in comprehending lipid-associated risk factors in patients with AMI, thereby enhancing clinical decision-making despite low AUC. Subsequent research should focus on larger sample sizes and integrating external validation datasets to enhance the assessment of the prediction efficacy of the model across varied patient demographics. The study was limited by cost that restricted the capacity for additional investigations and validations, especially in quantifying HDL-2b and HDL-3 levels via commercial laboratories. Further funding will be pursued for external validation data and for initiatives to broaden the study’s scope, including the validation of the model with independent cohorts. Finally, the limited ability to discern moderate categories in the multiclass classification task with the EBM model suggests that in future studies, Gensini score analysis may be more effective if the scores are transformed into binary variables.

Conclusions

HDL-3 provides superior predictive value for evaluating coronary artery stenosis severity in AMI patients compared to HDL-2b and other conventional lipid markers. The use of EBM can provide accurate delineation of HDL-2b and HDL-3 risk stratification because EBM has greater interpretability and clinical applicability. The concordance between EBM-derived cut-off points and guideline-recommended targets for LDL-C concentrations demonstrates the model’s robustness and its potential as a methodological benchmark for evaluating clinical markers. These findings support the integration of HDL subtype analysis into standard lipid profiling, which could lead to improved cardiovascular risk assessment and management. Further research should explore the longitudinal implications and broader applicability of these findings.

Supplementary material

Supplementary material is available at European Heart Journal – Digital Health.

Funding

Not available.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author upon reasonable request. Python code used in this study can be accessed at the GitHub address: https://github.com/tiaoguabi/HDL_multiclass.

Lead author biography

Bin Wang, MD, is a postdoctoral fellow at the School of Clinical Medicine, Tsinghua University. He received an MD degree in Clinical Research Methodology from Peking University Health Science Center. His research interests focus on machine learning.