We retrospectively enrolled hospitalized patients at Quzhou People’s Hospital from July 2021 to March 2025. The condition for positive sample collection is based on patients diagnosed as coronary atherosclerotic heart disease after discharge and whose length of stay is ≥ 2, excluding patients with cancer/tumor. A total of 19,690 eligible patients with CHD were included. The negative sample set comprised patients without a discharge diagnosis of cardiovascular-related diseases, hospitalized for ≥ 2 days, excluding patients with cancer/tumors. Ultimately, 17,765 eligible non-cardiovascular disease (non-CHD) patients were included.

Utilizing a healthy population as a control often causes the model to learn merely the generalized differences between ill patients and healthy individuals, rather than the specific pathological features that distinguish CHD from other diseases. Consequently, applying such a model to patients with existing comorbidities often results in unacceptable false positive rates, diminishing the clinical utility of the predicted results. To achieve the goal of disease screening within medical institutions and self-screening among individuals with underlying conditions, this research innovatively used other hospitalized patients with non-cardiovascular diseases as negative controls.

The dataset comprises three variable categories: demographic characteristics, blood biomarkers, and vital signs, all extracted from electronic health records (EHR). The basic personal information includes the patient’s gender, blood type, and age; lifestyle factors included smoking and drinking status; and comorbidities included diabetes and hypertension. Age was recorded at the time of treatment; smoking and drinking status were obtained from medical history records; and diabetes and hypertension status were derived from discharge diagnoses. Blood biomarkers, derived from initial admission tests, included complete blood counts (CBC), biochemical indicators (e.g., liver and kidney function), and D-dimer levels, among others. The vital sign information included the patient’s initial admission temperature, heart rate, respiratory rate, systolic blood pressure (SBP), and diastolic blood pressure (DBP), and SpO₂. Additionally, the maximum, minimum, and mean values of SBP, DBP, body temperature, respiratory rate and SpO₂ measured during hospitalization were recorded.

In laboratory testing, sample quality issues caused by hemolysis, instrument errors, or other factors can produce extreme outliers. As these outliers do not reflect true physiological or pathological states, rigorous detection and cleaning were performed on the blood biomarker data. We adopted a modified Z-score method to identify outliers, which is more robust to outliers (Kuo et al., 2024). M A D = m e d i a n X i − X m Z − s c o r e = 0.6745 × X i − X m   /   M A D

Among them, X i is the sample feature value, X m is the median of the sample feature value, and MAD is the median absolute deviation. Values with a Z-score > 3.5 were identified as outliers and replaced with null values (NaN). Further screening was conducted on samples with missing values below 40%, while retaining samples with more valid data. Consequently, 12 CHD and 7 non-CHD samples were removed.

The partial continuous value feature names, abbreviations, units, distribution descriptions (mean, standard deviation), and missing rates on the positive and negative sample sets are shown in the Table 1. Given the high dimensionality of the dataset, the complete feature table is included in the Supplementary Material.

Generative adversarial imputation Nets (GAIN) represent a data imputation method based on generative adversarial networks (GANs) (Xu et al., 2025; Nayak et al., 2024). By leveraging adversarial training between a generator and a discriminator, GAIN learns the underlying data distribution to generate plausible imputed values. The architecture is illustrated in Figure 1. The generator, which serves as the core component for imputation, utilizes a three-layer fully connected neural network structure. The input consists of the original data tensor X d a t a containing missing values and the mask tensor M d a t a , where 1 denotes observed data and 0 denotes missing data. The discriminator is tasked with distinguishing between observed true values and imputed values produced by the generator; its network structure mirrors that of the generator. The discriminator input comprises a concatenated tensor of the imputed data X i m p u t e d and the hint vector H d a t a . The hint vector, derived from a randomly generated probability matrix and a mask tensor, provides auxiliary information to assist the discriminator in identifying missingness patterns. The generator loss is composed of a weighted adversarial loss G loss _ 1 and an MSE loss G loss _ 2 , where the adversarial loss is achieved by minimizing the discriminator’s recognition accuracy of the generated values, and the MSE loss constrains the generator to not destroy the original information at known data positions. Optimize the discriminative ability of the discriminator by calculating the classification loss D loss between the original real data and the generated imputed data. The entire model gradually learns the inherent distribution pattern of the data through a continuous adversarial game between the generator and discriminator, ultimately generating missing values that are close to the true distribution.

To comprehensively evaluate the model’s predictive capability and clinical utility, we employed multiple complementary metrics, including accuracy, precision, recall, F1-score and Area Under the Receiver Operating Characteristic Curve (AUC) (Chen et al., 2025a; Kumar et al., 2024; Rimal and Sharma, 2023; Chen et al., 2025b; Qiao et al., 2024). These are the core metrics for evaluating the performance of binary classification models, calculated based on four fundamental values in the confusion matrix: true positive cases (TP), false positive cases (FP), true negative cases (TN), and false negative cases (FN) (Zeng et al., 2025; Zulfiqar et al., 2024; Qiao et al., 2025; Xie et al., 2025; Wang et al., 2025; Wang et al., 2024). A C C = T P + T N   /   T P + T N + F P + F N P r e c i s i o n = T P   /   T P + F P R e c a l l = T P   /   T P + F N F 1 = 2 × P r e c i s i o n × R e c a l l   /   P r e c i s i o n + R e c a l l

Figure 2 showed the visualization results of positive and negative samples using t-distributed stochastic neighbor embedding (t-SNE), which was used to display the distribution patterns of CHD and non-CHD samples in a high-dimensional feature space, intuitively presenting the feature differences and clustering patterns of the two groups (Nollmann et al., 2024).

In the scatter plot, blue (non-CHD) samples form a core cluster and two independent small clusters, indicating that the characteristics of the non-CHD population have strong concentration. The red (CHD) samples are interspersed in the form of scattered dots within and at the edges of the blue clusters, with only mild clustering in local areas and a relatively scattered overall distribution. In the density contour map, the density contour of non-CHD samples covers most of the areas in the map, and the core area has a high density, further verifying the concentration of non-CHD population characteristics and the stability of subgroup structure. The density profile of CHD samples highly overlaps with non-CHD, with only weak independent trends in local areas, indicating that the characteristic boundaries between CHD and non-CHD are blurred and overlap is high. At the same time, the heterogeneity of CHD features, such as different disease courses, subtypes, and comorbidities leads to their scattered distribution.

Figure 3 showed the heatmap of feature missing rates for the CHD and non-CHD groups. The missing rates for most variables were similar in both groups, but significant differences existed in key clinical indicators. The CHD group had significantly higher missing rates for hs-cTnI (38.88%) and HbA1c (33.12%) than the non-CHD group (9.32% and 5.53%, respectively). This difference may reflect insufficient detection of these important diagnostic and monitoring indicators in patients with CHD in clinical practice. In contrast, the non-CHD group had slightly higher missing rates for indicators such as D-dimer (20.37%) and hs-CRP (26.16%). The missing rates for most routine laboratory indicators and vital signs remained below 10% in both groups, indicating relatively complete basic clinical data collection. However, key indicators with high missing rates require appropriate missing data processing strategies in subsequent analyses.

Although incomplete recorded data may be reasonable in clinical practice, the performance of machine learning algorithms is often affected by biased and incomplete data. Medical record data is extremely valuable for disease research. If partially missing samples are directly removed and models are constructed using non missing samples, although this approach is simple, it wastes a lot of available information.

This research comprehensively compared and analyzed the performance of traditional imputation algorithms (KNN, MICE) (Varol et al., 2025), deep learning autoencoder series (AE, DAE, VAE) (Gautier et al., 2024; Shi et al., 2024), and GAIN in medical record data. Due to the lack of real data references, the performance of the downstream tasks is generally taken as the standard. Table 2 compared the 5-fold cross validation (5-cv) performance of the imputation algorithms, and Table 3 compared its performance on the independent test set. The performance metrics of traditional imputation algorithms were significantly lower than those of deep learning methods, and they were limited to being unable to adapt to the complex nonlinear correlations between features in the data, resulting in insufficient expression ability in high-dimensional medical data scenarios.

As a basic autoencoder, AE achieved a 5-cv AUC of 0.9146 and an independent test AUC of 0.9095, which preliminarily demonstrated the modeling ability of deep learning for complex data. DAE enhanced robustness through noise reduction mechanism, further improving performance (5-cv AUC 0.9157, independent test AUC 0.9135), and had a higher tolerance for data noise. After introducing variational inference, VAE had better flexibility in distribution modeling, with a 5-cv AUC 0.9196 and independent testing AUC 0.9165, ranking among the top in multiple performance metrics. As a generative model, GAIN not only considered the distribution of individual features in adversarial learning, but also comprehensively considers the complex correlations between all other features. Its 5-cv AUC 0.9154 and independent test AUC 0.9156 were slightly lower than VAE.

The primary goal of imputation in medical research is to faithfully preserve the original data distribution and minimize bias. We further observed the fitting degree of each imputation algorithm on the data distribution, and selected hs-cTnI and Jaun features with high missing rates. The feature density curves before and after imputation were shown in Figures 4, 5. Compared to VAE, the data distribution after GAIN imputation has a higher degree of fit with the original data in terms of morphology, which can avoid additional bias caused by distribution offset due to imputation and ensure the authenticity and effectiveness of feature information in subsequent analysis. To strictly quantify this observation, we used Kullback-Leibler (KL) divergence and Kolmogorov-Smirnov (KS) test statistics for the imputed data of VAE and GAIN on independent test sets. A lower KL or KS value indicates a distribution closer to the ground truth., GAIN achieved the lowest mean KL divergence (0.158) and mean KS statistic (0.067) outperforming VAE (KL: 0.210; KS: 0.068). This statistical evidence demonstrates that GAIN is superior in capturing the complex underlying probability distribution of the real data, avoiding the distributional shifts often introduced by variational inference in VAEs. Consequently, considering both the robust downstream performance and the superior data fidelity, GAIN was selected as the optimal imputation algorithm for this research.

In this section, we conducted a performance comparison analysis of XGBoost, random forest, logistic regression, SVM, KNN, AdaBoost, and ANN algorithms. The prediction metrics of training set in Table 4 showed that XGBoost (AUC = 0.9184, Accuracy = 0.8345) and ANN (AUC = 0.9182, Accuracy = 0.8381) have the most outstanding comprehensive performance. In the test set prediction performance in Table 5, XGBoost exhibited excellent generalization stability, while ANN’s generalization ability is significantly insufficient. Figure 6 showed the comparison of ROC curves of different prediction algorithms on the training and testing sets, which intuitively proves that the XGBoost model had the strongest ability to distinguish positive and negative samples and excellent generalization.

In many research, derived indicators based on blood biomarkers have shown excellent performance. We have established 8 derived indicators according to the obtained blood markers, namely, plasma atherogenic index (AIP = log₁₀ [TG/HDL]), uric acid/high-density lipoprotein ratio (UHR = UA/(18 * HDL)), neutrophil/lymphocyte ratio (NLR = NE/LYM), platelet/lymphocyte ratio (PLR = PLT/LYM), monocyte/lymphocyte ratio (MLR = MO/LYM), and systemic immune inflammation index (SII = PLT × NE/LYM), systemic inflammatory response index (SIRI = NE × MO/LYM), and systemic inflammatory composite index (AISI = NE × MO × PLT/LYM) (Wu et al., 2023; E et al., 2025).

In this section, we focused on different types of features, such as basic information (BI), blood biomarkers (BB), vital signs information (VSI), derivative indicators (DI). Feature combinations analysis was conducted, and its performance on the test set is shown in Table 6. From the view of single category features, blood biomarkers (BB) demonstrated core prediction value, reflecting the direct correlation of blood biomarkers in the pathological mechanisms of CHD such as lipid metabolism and inflammatory response. In multi class feature combinations, the performance of three class feature fusion (BI+BB+VSI) reached its peak, with accuracy 0.8218 and AUC 0.9047 being the best among all combinations. The integration of basic information, blood biomarkers, and vital signs has constructed a complete CHD risk profile from three dimensions, clinical phenotype, biochemical mechanisms, and physiological status, maximizing the complementarity between features. Although the feature combination performance of DI is theoretically guaranteed, it does not exceed BI+BB+VSI. This may be because DI introduced redundant information, which slightly interferes with the model’s generalization. This also proved the prediction algorithm’s ability to mine the cross-complementarity of feature.

The highest mean absolute SHAP value of mean-RR in Figure 6 indicated that it has the most significant global influence on CHD prediction in the model, followed by age, hs-cTnI, and hypertension, which collectively constitute the core drivers of model decision-making. We further analyzed the SHAP dependency plots for key features in Figure 7. The mean-RR dependency plot showed that a low respiratory rate is weighted as a positive contributor to CHD risk. In hospitalized patients or underlying disease populations, shortness of breath is an extremely common non-specific symptom with various causes, such as pain, anemia, anxiety, etc. Through data-driven analysis, the model identified high RR as strongly correlated with non-CHD hospitalization causes. Consequently, a relatively lower RR served as a distinguishing signal for occult CHD within this specific patient population. Regarding age, SHAP values increased monotonically, confirming age as a robust risk factor. The observed plateauing effect in the elderly suggested a deceleration in risk accumulation, consistent with established clinical knowledge regarding the progression of coronary atherosclerosis. For hs-cTnI, when hs-cTnI exceeded 0.0024, the SHAP value rapidly turned positive and remained at a high level, even within the clinical normal reference range, indicating that an increase in hs-cTnI has significantly increased the risk of CHD. This highlighted the sensitivity of high-sensitivity troponin in early myocardial injury and risk prediction. The role of SHAP analysis is limited to reflecting specific behavioral patterns of the model and cannot be used to infer causal relationships. At the same time, the numerical values of its results will also vary with the distribution of data and the structure of the model.

However, the control group may include individuals with acute respiratory conditions, potentially introducing a confounding bias where elevated respiratory rates reflect the pathology of the control group rather than a direct risk factor for CHD. To address this concern and verify the model’s robustness, we conducted a sensitivity analysis by excluding all respiratory-related features (mean-RR, max-RR, min-RR, and 1st-RR) and retraining the XGBoost model. The results showed that while the AUC on the test set experienced a moderate decline from 0.9053 to 0.8693, it remained within a clinically excellent range. This performance retention confirmed that although respiratory rate significantly contributes to discrimination, the model’s prediction power is fundamentally driven by the comprehensive integration of multiparametric features, rather than solely relying on distinguishing respiratory-related anomalies in the control group.

To further explore feature interactions, we calculated the Spearman rank correlation coefficients for the top 20 features ranked by SHAP importance, as visualized in Figure 8. The absolute value of the correlation coefficient was represented by the radius of a circle. The larger the radius, the stronger the correlation. The color represented the direction of correlation, with red indicating positive correlation and blue indicating negative correlation. The “X” in the upper triangle indicated that the feature pair is not statistically significant, while the specific correlation coefficient values were labeled in the lower triangle. The diagonal represented the autocorrelation of the feature. In addition to RR series features, the correlation coefficients of lipid metabolism features Ch, ApoB, and LDL are 0.79, 0.77, and 0.76, respectively, reflecting their synergistic effects in the process of lipid transport and atherosclerosis. This type of strong correlation prompt required attention to the joint effect of feature groups when interpreting model decisions, rather than the independent contribution of a single feature. However, features without statistically significant associations, such as age and some metabolic indicators, smoke and LDL, had no statistically significant association at the SHAP level (P > 0.05), indicating that the weight allocation of these features by the model was relatively independent and can reduce the interference of multicollinearity on model stability.

To facilitate the clinical translation of this low-cost screening tool, a dual approach can be implemented: integration into in-hospital EHR systems to generate automated risk alerts for non-cardiology departments, and the development of mobile health applications to enable self-monitoring for individuals with underlying conditions. However, widespread deployment necessitates addressing critical barriers, including strict adherence to data privacy regulations, the necessity of dynamic model updating to counter concept drift, and the challenge of fostering clinician trust—which is partially mitigated by the SHAP interpretability framework employed in this research. Furthermore, the interpretation of findings must be tempered by the limitations of a retrospective, single-center design. While the current model incorporates age and comorbidities as features, rigorous validation in future multi-center prospective studies is required to ensure fairness, robustness, and equitable healthcare outcomes across diverse subpopulations.