AI approaches for predicting progression to acute coronary syndrome among stable coronary heart disease patients

Abstract

Coronary heart disease (CHD), including its acute manifestation, acute coronary syndrome (ACS), remains a leading cause of morbidity and mortality worldwide. Early identification of ACS is critical for reducing the global burden of CHD, yet traditional diagnostic methods are often invasive, costly, and time consuming. In this study, we introduce an AI-driven approach that utilizes electronic health records (EHR) to identify transitions from stable CHD to ACS, using a comprehensive dataset of 12,336 patient records from 131 medical institutions in Jiangsu Province, China. The proposed model, applying the T2G-Former to predict ACS in 12 months, demonstrated superior performance, with an area under the curve of 0.953 and a sensitivity of 0.814 on the test set. Model interpretability was supported by SHAP values, which clarified the contribution of individual clinical variables. Our findings highlight the potential of AI-enabled EHR analysis for scalable early ACS detection and clinical decision support.

Introduction

Acute coronary syndrome (ACS) carries a strikingly high mortality burden: 30-day case fatality rates range from 6% to 11% in developed countries and are higher in developing regions¹, with mortality continuing to rise at 1 year post-event. Even among survivors, ACS frequently results in persistent myocardial dysfunction, heart failure, and recurrent ischemic episodes. As the acute manifestation of established coronary heart disease (CHD), ACS often strikes without warning—irreversible myocyte injury may begin within the first hour of symptom onset, leaving clinicians only a narrow window for effective intervention². With CHD currently affecting more than 250 million individuals worldwide, it is neither practical nor cost-effective to continuously monitor every at-risk patient³. Consequently, it is both desirable and essential to develop robust probabilistic models that stratify patients with stable CHD, who are diagnosed with cardiovascular disease but without a recent acute event, by their risk of progression to ACS. Such models would enable timely, targeted preventive and therapeutic measures, thereby improving outcomes.

Current screening techniques are often constrained by either low diagnostic accuracy, as in the case of exercise electrocardiography stress testing, or by invasiveness and high costs, as seen in coronary angiography and even in advanced non-invasive modalities like myocardial perfusion imaging and coronary computed tomography angiography^4,5. While these specialized cardiac imaging techniques can benefit many patients, their use is inappropriate in patients with a low pre-test probability of ACS^6,7. Given the large at-risk population and the existing prevention and treatment options, the development of an efficient and effective clinical tool for ACS prediction could enhance patient management, improve outcomes, and alleviate healthcare burdens.

The advent of machine learning (ML) and Artificial Intelligence (AI) technologies in the medical field has significantly transformed the prediction and management of diseases^8–10. Recent advancements have demonstrated the potential of ML methods to enhance CHD diagnostic accuracy and efficiency. For instance, Bock et al. developed ML models that outperformed cardiologists in predicting stress-induced, functionally relevant CHD. By using eight static clinical variables and electrocardiogram signals from exercise stress testing, their models could potentially reduce unnecessary imaging procedures by 15–17% while improving diagnostic accuracy¹¹. Similarly, Cheng et al. validated bayesian network, logistic regression, random forest, neural network, and gradient boosting models for CHD prediction, achieving an area under the curve (AUC) of 0.846 and demonstrating high diagnostic performance¹². Petrazzini et al. also utilized clinical variables from electronic health records (EHRs) to assess CHD risk, further highlighting the efficacy of ML methods in this domain¹³.

While existing studies have demonstrated the utility of ML techniques in CHD diagnosis and risk prediction using static clinical variables, imaging data, or stress-test data, few have specifically focused on developing temporally aware models to predict progression from stable CHD to ACS using longitudinal, multidimensional EHR data. With the widespread adoption of EHRs, this represents a timely opportunity to harness AI methods, which are uniquely capable of integrating heterogeneous data and capturing dynamic patient trajectories. Although challenges such as missing data, coding inconsistencies, and institutional heterogeneity remain, advances in preprocessing, harmonization, and interpretability techniques increasingly mitigate these issues, highlighting the strong potential of AI-driven models to transform longitudinal EHR data into clinically actionable insights.

This study aims to develop an accurate and efficient model for predicting progression to ACS in patients with stable CHD. By incorporating longitudinal records, comorbidities, medications, and other relevant clinical information that may not be captured in traditional diagnostic methods, the model will provide clinicians with a dynamic, scalable tool for early risk identification. Such a model can facilitate timely interventions, optimize resource allocation, and ultimately reduce the burden of preventable ACS events. Furthermore, elucidating key predictive features may uncover novel biomarkers or risk modifiers, advancing mechanistic understanding of CHD progression. In an era where personalized medicine is paramount, this work has the potential to translate AI innovations into clinically actionable insights for the management of stable CHD.

Results

Datasets

Between March 1, 2020, and February 28, 2023, a total of 268,876 CHD patient records were collected across 165 medical institutions and consolidated by the National Heart Disease Database Consortium, covering more than 200 clinical variables. Since the EHR data, including sociodemographic characteristics, healthcare utilization measures, disease diagnoses, and laboratory test results, are heterogeneous across institutions, we converted them into a structured tabular format to facilitate analysis. To deal with the missing records, quality control procedures were then applied, excluding variables and individuals with more than 50% missing data. Supplementary Table 2 summarizes the missingness of each clinical feature prior to imputation. Missing values were subsequently imputed using multiple imputation by chained equation (MICE) on the longitudinal data, thereby preserving individual-level trends without discarding any records¹⁴. The final refined dataset consisted of 34 clinical variables across 12,336 records from 131 institutions. Among these patients with stable CHD, 891 progressed to ACS within 6 months, 1227 within 12 months, and 1407 within 24 months (Fig. 1b).

Fig. 1. The workflow for data collection and model development.

a Use Python to organize EHR data into tabular data. During data cleansing, variables and individuals with more than 50% missing values were excluded. b Study individuals. c Develop AI-based predictive models using tabular data. NHBDC National Healthcare Big Data (East) Center, CHD coronary heart disease, ACS acute coronary syndrome, SVM support vector machine, XGBoost Extreme Gradient Boosting.

To explore the distributions of clinical variables, we compared patients with stable CHD and those who progressed to ACS within 24 months (Supplementary Fig. 1). Two-sided Mann–Whitney U tests were performed for each variable, revealing statistically significant differences for most biomarkers (p < 0.05), such as Cholinesterase and Low-Density Lipoprotein Cholesterol (Supplementary Table 4). However, in down-sampling sensitivity analyses (Supplementary Fig. 2), p values decreased monotonically with larger samples, whereas ROC-AUC, effect sizes (Cohen’s d), and distribution-overlap measures (OVL) remained essentially constant, confirming that the apparent significance is primarily sample-size-driven rather than clinically meaningful. Notably, Thyroid Stimulating Hormone (TSH) and Glucose (Glu) showed no statistically significant difference between groups.

AI-based predictive model

In this study, we developed and validated a deep-learning-based predictive model, which consisted of two stages: (1) construction of embeddings; (2) prediction of ACS probabilities within 12 months.

Because EHR features often exhibit complex correlations and structure, we built a sparse feature relation graph by thresholding pairwise associations and retaining only significant edges. We then applied T2G-Former¹⁵ to this pruned graph to extract patient-level embeddings. Additionally, we assumed that all records are independent to simplify the embedding process. At the second stage, we fed embeddings into a multilayer perceptron (MLP) that outputs the probability of ACS within 12 months. To mitigate the extreme class imbalance between stable CHD and ACS patients, we optimized a weighted binary-cross-entropy (weighted BCE) loss in which the positive (ACS) class is assigned a larger weight¹⁶. The T2G-Former and MLP head were trained end-to-end by minimizing this loss, yielding our final predictive model for 12-month ACS risk.

Comparison of different methods

After processing all clinical variables, we evaluated a range of predictive models to estimate patient-level ACS risk, including traditional ML methods such as CatBoost (CB), LightGBM (LGBM), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost)^17–21. To compare deep-learning approaches on tabular data, we additionally assessed another transformer-based architecture, specifically FT-Transformer²², as shown in Fig. 1c.

By comparing different models (Fig. 2 and Tables 1–3), T2G-Former consistently outperforms alternatives across multiple metrics. It achieves the highest F1 scores and AUC values across all scenarios, and the highest sensitivity and precision in most settings. While traditional machine-learning models such as LightGBM and XGBoost also produced high AUCs in the experiments, clinical decision-making for a life-threatening event like ACS prioritizes minimizing missed diagnosis (false negatives), and hence sensitivity is a more clinically relevant performance measure. In details, F1 scores are 0.6664 ± 0.0126 (6-month model, mean ± s.d.), 0.7567 ± 0.0202 (12-month model), 0.8855 ± 0.0083 (24-month model), AUC values are 0.9331 ± 0.0038 (6-month model), 0.9528 ± 0.0039 (12-month model), 0.9821 ± 0.0012 (24-month model), sensitivity values are 0.7355 ± 0.0280 (6-month model), 0.8139 ± 0.0103 (12-month model), 0.8635 ± 0.0143 (24-month model), precision values are 0.6105 ± 0.0248 (6-month model), 0.7084 ± 0.0391 (12-month model), 0.9095 ± 0.0265 (24-month model).

Fig. 2. Receiver-operating characteristic curves for all models.

a, b are the training and test results for the 6-month model. c, d are the training and test results for the 12-month model. e, f are the training and test results for the 24-month model.

Table 1.

Classification performance of the ML and DL models based on the dataset of CHD patients progressed to ACS within 6 months (weighted BCE)

	Model	Precision	Sensitivity	F1-core	Acc	AUC
Validation
	CatBoot	0.6216 ± 0.0685	0.2756 ± 0.0818	0.3763 ± 0.0877	0.9355 ± 0.0045	0.9195 ± 0.0103
	LGBM	0.6363 ± 0.0348	0.3893 ± 0.0060	0.4803 ± 0.0537	0.9396 ± 0.0034	0.9327 ± 0.0088
	RandomForest	0.6041 ± 0.0585	0.2145 ± 0.0676	0.3101 ± 0.0796	0.9328 ± 0.0032	0.9078 ± 0.0124
	Xgboost	0.6326 ± 0.0435	0.3671 ± 0.0876	0.4595 ± 0.0808	0.9388 ± 0.0044	0.9325 ± 0.0094
	SVM	0.4284 ± 0.0287	0.7768 ± 0.0341	0.5513 ± 0.0239	0.9081 ± 0.0094	0.9147 ± 0.0098
	FTtransformer	0.5944 ± 0.0359	0.5862 ± 0.0997	0.5862 ± 0.0439	0.9404 ± 0.0036	0.9323 ± 0.0077
	T2GFormer	0.5821 ± 0.0206	0.7022 ± 0.0283	0.6358 ± 0.0080	0.9417 ± 0.0028	0.9338 ± 0.0030
Test
	CatBoot	0.6318 ± 0.0682	0.2603 ± 0.0518	0.3651 ± 0.0548	0.9354 ± 0.0035	0.9217 ± 0.0078
	LGBM	0.6242 ± 0.0422	0.3497 ± 0.0850	0.4416 ± 0.0715	0.9377 ± 0.0032	0.9317 ± 0.0082
	RandomForest	0.6079 ± 0.0512	0.2079 ± 0.0663	0.3021 ± 0.0766	0.9328 ± 0.0021	0.9100 ± 0.0133
	Xgboost	0.6200 ± 0.0586	0.3380 ± 0.0789	0.4312 ± 0.0674	0.9369 ± 0.0043	0.9327 ± 0.0059
	SVM	0.4192 ± 0.0171	0.7814 ± 0.0291	0.5452 ± 0.0152	0.9059 ± 0.0059	0.9196 ± 0.0080
	FTtransformer	0.5606 ± 0.0372	0.5370 ± 0.1241	0.5398 ± 0.0667	0.9357 ± 0.0044	0.9217 ± 0.0121
	T2GFormer	0.6105 ± 0.0248	0.7355 ± 0.0280	0.6664 ± 0.0126	0.9469 ± 0.0033	0.9331 ± 0.0038

Bold values indicate the best performance for each metric.

Table 3.

Classification performance of the ML and DL models based on the dataset of CHD patients progressed to ACS within 24 months (weighted BCE)

	Model	Precision	Sensitivity	F1-core	Acc	AUC
Validation
	CatBoot	0.9682 ± 0.0119	0.5641 ± 0.0503	0.7116 ± 0.0406	0.9482 ± 0.0057	0.9372 ± 0.0094
	LGBM	0.9737 ± 0.0126	0.7152 ± 0.0367	0.8240 ± 0.0231	0.9613 ± 0.0036	0.9601 ± 0.0077
	RandomForest	0.9938 ± 0.0053	0.5944 ± 0.0463	0.7429 ± 0.0369	0.9534 ± 0.0053	0.9313 ± 0.0132
	Xgboost	0.9730 ± 0.0139	0.6996 ± 0.0502	0.8129 ± 0.0348	0.9635 ± 0.0055	0.9586 ± 0.0072
	SVM	0.6422 ± 0.0420	0.8093 ± 0.0313	0.7146 ± 0.0229	0.9260 ± 0.0092	0.9433 ± 0.0092
	FTtransformer	0.8569 ± 0.0583	0.7804 ± 0.0377	0.8150 ± 0.0296	0.9595 ± 0.0073	0.9646 ± 0.0080
	T2GFormer	0.9234 ± 0.0259	0.8625 ± 0.0114	0.8916 ± 0.0098	0.9637 ± 0.0039	0.9860 ± 0.0011
Test
	CatBoot	0.9705 ± 0.0130	0.5827 ± 0.0513	0.7269 ± 0.0406	0.9505 ± 0.0059	0.9396 ± 0.0165
	LGBM	0.9731 ± 0.0178	0.7242 ± 0.0566	0.8293 ± 0.0377	0.9593 ± 0.0067	0.9650 ± 0.0110
	RandomForest	0.9956 ± 0.0055	0.5985 ± 0.0655	0.7455 ± 0.0496	0.9540 ± 0.0074	0.9414 ± 0.0187
	Xgboost	0.9655 ± 0.0212	0.7061 ± 0.0545	0.8146 ± 0.0376	0.9606 ± 0.0065	0.9623 ± 0.0119
	SVM	0.6634 ± 0.0566	0.8025 ± 0.0466	0.7258 ± 0.0498	0.9307 ± 0.0135	0.9437 ± 0.0154
	FTtransformer	0.8446 ± 0.0678	0.7785 ± 0.0549	0.8081 ± 0.0459	0.9578 ± 0.0108	0.9588 ± 0.0120
	T2GFormer	0.9095 ± 0.0265	0.8635 ± 0.0143	0.8855 ± 0.0083	0.9615 ± 0.0034	0.9821 ± 0.0012

Bold values indicate the best performance for each metric.

Table 2.

Classification performance of the ML and DL models based on the dataset of CHD patients progressed to ACS within 12 months (weighted BCE)

	Model	Precision	Sensitivity	F1-core	Acc	AUC
Validation
	CatBoot	0.8477 ± 0.0323	0.4596 ± 0.0458	0.5941 ± 0.0368	0.9380 ± 0.0036	0.9319 ± 0.0085
	LGBM	0.8533 ± 0.0385	0.5923 ± 0.0859	0.6949 ± 0.0586	0.9491 ± 0.0068	0.9510 ± 0.0064
	RandomForest	0.8631 ± 0.0413	0.4711 ± 0.0751	0.6044 ± 0.0656	0.9398 ± 0.0059	0.9212 ± 0.0116
	Xgboost	0.8475 ± 0.0352	0.5578 ± 0.0833	0.6688 ± 0.0632	0.9460 ± 0.0070	0.9494 ± 0.0070
	SVM	0.5679 ± 0.0170	0.7972 ± 0.0259	0.6628 ± 0.0104	0.9195 ± 0.0039	0.9332 ± 0.0056
	FTtransformer	0.7903 ± 0.0358	0.6990 ± 0.0417	0.7409 ± 0.0302	0.9515 ± 0.0053	0.9504 ± 0.0058
	T2GFormer	0.7280 ± 0.0438	0.8006 ± 0.0092	0.7616 ± 0.0222	0.9562 ± 0.0036	0.9525 ± 0.0022
Test
	CatBoot	0.8461 ± 0.0312	0.4677 ± 0.0387	0.6010 ± 0.0306	0.9386 ± 0.0031	0.9367 ± 0.0068
	LGBM	0.8511 ± 0.0337	0.6173 ± 0.0564	0.7138 ± 0.0363	0.9511 ± 0.0048	0.9524 ± 0.0050
	RandomForest	0.8591 ± 0.0280	0.4867 ± 0.0606	0.6184 ± 0.0470	0.9409 ± 0.0044	0.9262 ± 0.0092
	Xgboost	0.8482 ± 0.0298	0.5837 ± 0.0666	0.6889 ± 0.0448	0.9471 ± 0.0051	0.9518 ± 0.0053
	SVM	0.5776 ± 0.0331	0.7929 ± 0.0183	0.6677 ± 0.0233	0.9214 ± 0.0080	0.9348 ± 0.0069
	FTtransformer	0.7860 ± 0.0461	0.6898 ± 0.0369	0.7336 ± 0.0291	0.9503 ± 0.0057	0.9472 ± 0.0082
	T2GFormer	0.7084 ± 0.0391	0.8139 ± 0.0103	0.7567 ± 0.0202	0.9479 ± 0.0061	0.9528 ± 0.0039

Bold values indicate the best performance for each metric.

Although SVM achieves comparable or slightly higher sensitivity (0.7929 ± 0.0183 for the 12-month model and 0.7814 ± 0.0291 for the 6-month model), its precision is considerably lower, which are 0.4192 ± 0.0171 and 0.5776 ± 0.0331, respectively, limiting its practical utility.

We also compared models trained with weighted and unweighted BCE losses. Unweighted results for the 6-, 12-, and 24-month prediction windows appear in Supplementary Tables 5–7. Overall, both loss functions yield similar performance, and the weighted BCE slightly outperforms its unweighted counterpart at the 12- and 24-month windows. In terms of training efficiency, T2G-Former with weighted BCE required on average 134.5 epochs (12-month), 122.6 epochs (24-month), and 158.3 epochs (6-month), whereas those with unweighted BCE converged after 127.3 epochs (12-month), 154.0 epochs (24-month), and 149.6 epochs (6-month). The training losses are presented in Supplementary Fig. 3, which illustrates the convergence behavior of the models across different prediction windows and loss functions.

To verify that the observed effects are not artifacts of the imputation procedure, we performed a sensitivity analysis using an alternative k-nearest neighbors (k-NN) imputation method alongside MICE and found performance remained strong across all prediction windows (Supplementary Tables 8–10): test AUCs (mean ± SE) for k-NN vs MICE were 0.9245 ± 0.0033 vs 0.9331 ± 0.0038 at 6 months, 0.9396 ± 0.0045 vs 0.9528 ± 0.0039 at 12 months, and 0.9642 ± 0.0033 vs 0.9821 ± 0.0012 at 24 months. In addition, we quantified empirical null distributions of AUC via label-permutation tests (Supplementary Fig. 4), using train-label permutation with retraining (number of permutations = 200) and test-label permutation without retraining (metric-only, number of permutations = 2000). The resulting nulls were centered near 0.50 with 95% ranges approximately 0.41–0.59, and the observed validation and test AUCs lay above the 97.5% quantiles of these nulls. Collectively, the concordant AUCs under k-NN versus MICE and the null AUC distributions from train-retrain and metric-only permutation tests indicate that T2G-Former’s predictive gains are robust and not attributable to chance or analysis artifacts.

Comparison of models across different time windows

We evaluated models for predicting ACS risk within the next 6, 12, and 24 months, with the corresponding ROC curves shown in Fig. 2 and other detailed evaluation results summarized in Tables 1–3. Model performance improved as the prediction window lengthened. Specifically, the 24-month model attained a sensitivity of 0.8635 ± 0.0143, precision of 0.9095 ± 0.0265, F1 score of 0.8855 ± 0.0083, and AUC of 0.9821 ± 0.0012. The 12-month model exhibited a modest decline, achieving an AUC of 0.9528 ± 0.0039 and a sensitivity of 0.8139 ± 0.0103, which are comparable to those of the 24-month model. Although the 6-month model also demonstrated a respectable AUC (0.9331 ± 0.0038), its sensitivity (0.7355 ± 0.0280) was relatively lower.

When predictive performances are comparable, shorter-term models offer distinct clinical advantages. They enable clinicians to identify high-risk patients within a nearer follow-up window, facilitating prompt therapeutic adjustments and closer monitoring to prevent adverse events. In addition, risk stratification over a shorter interval allows more efficient resource allocation, such as scheduling follow-up tests and consultations, and reinforces patient adherence by conveying a clearer sense of urgency. Accordingly, the 12-month model is recommended for clinical practice, as it provides earlier risk identification than the 24-month model while maintaining comparable performance.

Model interpretability

To help cardiologists better understand the functioning of a prediction model, it is essential to provide interpretability through post-hoc explanations. In our approach, we employed SHapley Additive exPlanations (SHAP) values, a method based on game theory, to explain the predictions of our T2G-Former models. The corresponding values are presented in Fig. 3.

Fig. 3. Top 20 features ranked by SHAP values from the prediction model across three prediction windows.

a 6 months, b 12 months, and c 24 months. Venn diagrams displaying overlapping SHAP-ranked features among three prediction windows: d top 5, e top 10, and f top 20 features. Abbreviations of clinical features are defined in Supplementary Table 1.

SHAP analysis consistently identified five key features across all three models, which have top SHAP values—total cholesterol (TC), Alanine aminotransferase (ALT), free thyroxine (FT4), total bilirubin (TBIL), and lymphocyte percentage (LYM%), all of which are supported by existing clinical studies. In clinical practice, elevated TC is a well-established risk factor for atherosclerosis and ACS²³. This finding is consistent with the SHAP summary plots, where higher TC values are associated with increased predicted risk of ACS. ALT often transiently rises during the acute phase of ACS, which likely reflects hepatic ischemia or congestion. Paradoxically, a relatively “low-normal” ALT level at admission has also been linked to worse in-hospital and long-term mortality among ACS patients, perhaps as a surrogate for underlying frailty, malnutrition, or sarcopenia²⁴. Thyroid function abnormalities, such as a low FT4, are frequently observed in ACS and correlate with larger infarct size, worse left ventricular dysfunction, and higher major adverse cardiovascular events (MACE) rates²⁵. TBIL, an endogenous antioxidant, is associated with ACS risk: mildly elevated TBIL appears protective against plaque oxidation and subsequent ACS, whereas very low TBIL, indicating poor antioxidant reserve, portends higher ACS incidence and worse post-ACS outcomes²⁶. As part of the white blood cell differential, LYM% reflects the body’s inflammatory stress response and immune function. Patients with ACS often exhibit stress-induced leukocytosis accompanied by a relative lymphopenia, and hence a reduced LYM%^27,28.

Aside from the five core predictors, the models rank other features differently: plasma fibrinogen (FIB) emerges as another important predictor in the 6-month and 12-month models, and C-reactive protein (CRP) is crucial in the 24-month model. This pattern likely reflects the fact that FIB and CRP each provide prognostic value at different follow-up time points. An elevated FIB level at admission often signals a hypercoagulable state in the acute phase and is closely associated with MACE occurring within 6 months after percutaneous coronary intervention or the acute episode²⁹. As a systemic inflammation marker, CRP offers strong prognostic information for long-term post-ACS outcomes^30–32, making it a critical feature in the 24-month model.

As shown in Supplementary Table 4 and Fig. 3, for all three models, the top-ranked features identified by SHAP analysis were all statistically different between stable CHD patients and those who developed ACS. Among the 34 variables, only TSH (P = 0.26) and Glu (P = 0.109) did not differ significantly between groups. Nevertheless, SHAP analysis ranked TSH consistently within the top-20 predictors across the 6-, 12-, and 24-month windows, and Glu within the top-20 in the 6-month window only, indicating conditional predictive value beyond univariate comparisons.

Single-feature partial dependence plots (PDPs) revealed clear nonlinearity: Glu displayed a U-shaped association with the lowest risk near normoglycemia (Supplementary Fig. 5a), whereas TSH showed a threshold-like rise in risk once exceeding the cohort mean, a pattern reproduced in the 6-, 12-, and 24-month models (Supplementary Fig. 5b–d). Two-way PDPs further revealed clinically coherent interactions (Supplementary Figs. 6–9). For Glu & FT4 and Glu & BUN, risk increased with higher glucose and was modulated by the partner marker, which forms a low-risk band at lower FT4 and a mid-range trough for BUN with higher risk at both ends. For TSH paired with other variables (e.g., LDL-C, CRP), the joint surfaces were dominated by TSH gradient with modest modulation by partner variables, consistent with the threshold-like 1D PDPs. Collectively, these findings reconcile the lack of univariate differences for TSH and Glu with their SHAP-identified importance by demonstrating non-linear and context-dependent effects.

Discussion

CHD and ACS remain the leading causes of morbidity and mortality worldwide. Early identification of individuals at high risk of progressing from stable CHD to ACS is critical for implementing timely interventions and reducing the overall burden of CHD; however, this task poses a significant clinical challenge. In this study, we developed and evaluated EHR-based models to predict this progression. Our findings demonstrate that these AI-driven models to enhance the diagnostic process, particularly in distinguishing between stable CHD and ACS, with implications for improving patient outcomes and optimizing healthcare resources. Leveraging routinely collected EHR records, these models have the potential to be used as a clinical decision support tool and can be easily integrated into existing EHR systems.

Although many models have been developed for CHD prediction, the existing methods have limited prediction accuracy^11,33,34. The proposed DL-based model has demonstrated superior performance across multiple evaluation metrics, including AUC and F1 scores, especially in the face of imbalanced medical data³⁵. By integrating the weighted BCE loss that assigns greater weight to the underrepresented ACS class, we further enhanced discrimination and reduced misclassification. Such improvement underscores the importance of addressing class imbalance in medical diagnostics, where both false positives and false negatives can have significant consequences. Given that ACS triage must avoid missed high-risk cases, improved sensitivity without sacrificing overall discrimination warrants the model’s increased complexity. Although training is more computationally demanding, once deployed, the model delivers low-latency predictions within existing EHR systems, underscoring its clinical feasibility for decision support.

Moreover, our study emphasizes the value of model interpretability, particularly in a clinical setting where decision transparency is essential for adoption. By utilizing SHAP values, we were able to provide clear insights into the contribution of individual variables to the model’s predictions, thereby aligning the model’s decision-making process with established clinical knowledge. This interpretability allows clinicians to understand not only which patients are at elevated risk of progressing from stable CHD to ACS, but also why the model arrived at that conclusion. Such information can facilitate personalized clinical decision-making. For instance, patients with elevated TC, ALT, FT4, TBIL, or LYM% may warrant closer monitoring to detect early signs of deterioration. Furthermore, by identifying modifiable risk factors, such as lipid abnormalities or elevated inflammatory markers, clinicians can implement tailored interventions, including intensified lipid-lowering regimens or targeted anti-inflammatory therapies. Collectively, these findings could encourage clinicians to incorporate both established and less emphasized biomarkers into their routine risk assessments, particularly for medium- to long-term ACS prevention strategies. While these associations cannot be interpreted as causal, they remain biologically plausible and clinically meaningful. By offering transparent, data-driven explanations, our approach fosters reliability among healthcare professionals and facilitates seamless integration of AI-driven tools into routine clinical workflows, ensuring that the models remain practical and relevant across different healthcare settings.

Beyond comparisons with clinical practice, we also benchmarked our feature design against recent computational and biomedical AI studies. Chen et al. developed a meta-prediction model for 10-year coronary artery disease risk by integrating polygenic risk scores, measured biomarkers, and meta-features³⁶. Their leading predictors—such as Lp(a), PLT, WBC, SBP, cholesterol/HDL ratio, HbA1c, apolipoprotein B, and cystatin C (the latter corresponding to creatinine in our dataset), overlapped with several of our laboratory and vital-sign covariates. In contrast, the MAARS framework for sudden arrhythmic death in hypertrophic cardiomyopathy integrates EHR-level demographics and stress-test vitals with echocardiographic and late gadolinium enhancement-cardiac magnetic resonance imaging, emphasizing structural and electrophysiologic features rather than detailed laboratory measurements³⁷. Our work complements these approaches by focusing on a hospital EHR-derived laboratory profile, including coagulation, inflammatory, and myocardial injury markers, tailored to forecasting ACS progression in the near term, a clinically important but underexplored prediction task in the current literature.

Our study focuses on EHR data, a readily available and routinely collected source, which underscores the practical applicability of our models in real-world clinical settings. By leveraging the rich information contained in EHRs, our models were able to achieve high diagnostic accuracy without the need for invasive or costly procedures, making them suitable for widespread implementation, particularly in resource-limited settings. At the same time, we recognize that large-scale EHR data present inherent challenges, as they are primarily designed for clinical care rather than research. This creates several difficulties for secondary analysis, including heterogeneous data entry and coding practices across hospitals, irregular laboratory testing that results in substantial missingness, and complex correlations among clinical variables that may introduce multicollinearity, noise, unmeasured confounding, and selection bias. To mitigate these issues, we implemented rigorous quality control, applied MICEs to address missing values while preserving longitudinal trends, and constructed a sparse feature relation graph within the T2G-Former framework to capture meaningful dependencies while reducing redundancy. These strategies enabled us to harness the strengths of “big data” EHRs while minimizing their limitations.

Our study also has limitations. First, although our dataset was derived from a single region in China, it included 131 tertiary hospitals across the province. As referral centers, these hospitals manage not only local urban patients but also complex cases referred from secondary and primary care facilities, thereby capturing a wide spectrum of demographic and clinical characteristics. This diversity provides the dataset with a reasonable degree of representativeness, which supports the potential generalizability of our modeling framework. Nevertheless, differences in population genetics, healthcare practices, and clinical environments across regions may still impact model performance. External validation in geographically and ethnically diverse cohorts will therefore be essential to confirm robustness and ensure broad clinical applicability. Second, the retrospective nature of our data collection introduces inherent biases, and the accuracy and completeness of EHR data may vary across sites. We mitigated these issues through quality control, variable filtering, patient-level longitudinal imputation, and MICE, but residual uncertainty cannot be eliminated. Third, while we applied multiple ML and DL models to predict ACS from stable CHD, all were trained and evaluated on historical data. Their performance in real-time, prospective clinical settings remains untested. Given the evolving nature of CHD and ACS, as well as ongoing advancements in medical practice, our models will require periodic retraining and external validation to maintain accuracy and clinical relevance.

In conclusion, our study demonstrates the feasibility and effectiveness of using AI models to predict ACS from EHR data. The models developed offer a promising tool for early detection and management of these conditions, with the potential to significantly reduce the burden on healthcare systems. Future research should focus on further validating these models in different populations and clinical settings, as well as exploring their integration into clinical decision support systems to enhance their impact on patient care. Ultimately, achieving optimal predictive performance and clinical acceptance will likely require a synergistic approach that combines AI-driven predictions with expert clinical judgment.

Methods

Data acquisition

In Jiangsu, the Jiangsu Commission of Health has established the National Healthcare Big Data (East) Center (NHBDC) to facilitate the secure sharing of EHR data across the province. NHBDC provides anonymized data encompassing a wide range of patient information collected from all public hospitals and clinics within Jiangsu Province. The EHR system offers an integrated, longitudinal view of patients’ health status and clinical outcomes, including comprehensive medication and laboratory records, hospitalization details, residential area linked to poverty index, health service utilization, comorbidities, and procedural data. We extracted a retrospective dataset from NHBDC, which includes patients with any inpatient admissions or outpatient attendances from March 1, 2020, to February 28, 2023. The data used was originally collected for routine patient management, requiring no additional data input for this study. Personal information was removed during the data analysis process to ensure privacy and confidentiality.

Data prepossessing

The data collected include sociodemographic data, residential areas (linked to average income as an indicator of social deprivation and poverty), and healthcare resource utilization (such as inpatient admissions, transfers and discharges, outpatient admissions, clinic visits, consultations with allied healthcare professionals, and emergency room visits). Additionally, we extracted disease diagnoses from International Classification of Diseases, Tenth Revision (ICD-10) codes, medication dispensing information, and laboratory data, including hematology, renal, and liver function tests, and glycemic and lipid indexes. To ensure data quality, variables with more than 50% missing values and individuals with more than 50% missing values were excluded.

Since not all patients underwent every test at each visit, missing values occur frequently in the dataset. Excluding all incomplete records would greatly diminish the sample size. To address this, we leveraged the longitudinal structure by imputing any missing test result with the mean of the corresponding patient’s available results across other visits. This preserves cohort size while respecting individual patient trajectories in the data. The remaining missing values were imputed using MICEs. Post-imputation, variables with a pairwise Pearson correlation coefficient greater than 0.90 were removed to prevent multicollinearity (Fig. 4).

Fig. 4. Heatmap of correlations among clinical variables.

Post-imputation, variables with a Pearson correlation coefficient greater than 0.90 were removed to prevent multicollinearity. All pairwise values were below 0.9; therefore, no variables were removed, and 34 variables remain. Abbreviations of clinical features are defined in Supplementary Table 1.

Patients with CHD were identified using ICD-10 codes (I20-I25) for coronary artery disease extracted from longitudinal EHR records. Stable CHD and ACS cases were further differentiated through specific subcodes (e.g., I20.8 for stable angina and I21.9 for unspecified acute myocardial infarction), ensuring alignment with clinical diagnostic criteria.

T2G-Former

T2G-Former is used to capture the interactions between features in the tabular data. By converting the tabular data into a graph, we can then apply a multi-head attention mechanism to extract embeddings of each patient record.

The framework consists of three steps: (1) Feature Relation Graph Estimation, (2) Graph-Guided Transformer Encoding, and (3) Cross-Level Readout and Pooling.

In details, using the fully imputed, flattened tabular matrix $X\in {\mathbb{R}}^{N\times p}$, compute the Pearson’s correlation for each pair of variables. By setting the threshold as 0.5, we can build an unweighted adjacency matrix $A\in {\mathbb{R}}^{\mathrm{p}\times p}$, by keeping the significant edges. Then we have a sparse undirected feature relation graph $G=\left(V,E\right)$ where $V=\{1\dots p\}$ are feature-nodes and $E=\{\left(i,j\right):A_{\mathit{ij}}=1\}$.

We then project each feature $j$ into an initial vector $h_j^{\left(0\right)}\in {\mathbb{R}}^d$, where the dimension d is set to 112. Set the number of layers $L$ to be 3. At layer $l$, for each node $i$ and each head $h$, compute attention only over its graph neighbors, then update $h_i^{\left(l+1\right)}$ by LayerNorm. Repeat the above for $l=0,\dots ,L-1$, to capture higher-order feature interactions, and the node can be then represented as $\{h_j^{\left(1\right)},\dots ,h_j^{\left(L\right)}\}.$

Lastly, for each feature node $j$, concatenate its 3-layer outputs into $\widetilde{h_j}=\left[h_j^{\left(1\right)};h_j^{\left(2\right)};h_j^{\left(3\right)}\right]$. Aggregate over all $p$ nodes to obtain a single patient embedding $H=\frac{1}{p}{\sum }_{j=1}^p{h}_j$. Then it would output a fixed-length vector $H$ summarizing all significant feature interactions, ready for downstream risk prediction.

This embedding construction process allowed the model to learn meaningful dependencies among heterogeneous EHR variables and to represent each patient with a fixed-length, clinically informative embedding suitable for downstream prediction.

Binary cross entropy and weighted binary cross entropy

BCE is a loss function for binary classification problems, quantifying the discrepancy between the true class labels $y_i\in \left\{\mathrm{0,1}\right\}$ and the predicted probabilities $p_i$. Mathematically, for a sample of size $N$, it is defined as

$$\mathcal{L}=-\frac{1}{N}\sum _{i=1}^N\left[y_i\log \left(p_i\right)+\left(1-y_i\right)\log \left(1-p_i\right)\right]$$

To address class imbalance, weighted BCE is proposed via introducing class-specific weights $w^{+}$ and $w^{-}$ for the positive class ($y=1$) and the negative class ($y=0$), yielding the modified loss function

$${\mathcal{L}}_w\left(p,y\right)=-\frac{1}{N}\sum _{i=1}^N\left[{w^{+}y}_i\log \left(p_i\right)+w^{-}\left(1-y_i\right)\log \left(1-p_i\right)\right]$$

Here weights are usually inversely proportional to the class sizes. A common choice is to set

$$w^{+}=\frac{N^{-}}{N^{+}+N^{-}}\mathit{and}{w}^{-}=\frac{N^{+}}{N^{+}+N^{-}},$$

where $N^{+}$ and $N^{-}$ are the sizes of positive and negative samples, respectively.

Cross-validation strategy

To ensure fair and robust model evaluation, we applied the same cross-validation protocol to all ML and transformer-based models. Specifically, a fixed random seed and a 5-fold stratified cross-validation were used to split the training-validation and test set, ensuring that the class distribution was preserved in each fold. The dataset was randomly split into a training-validation set (80%) and a test set (20%). The training-validation set was further partitioned, again in an 80/20 ratio, into training and validation subsets, with the former used to fit model parameters and the latter to optimize hyperparameters. The final hyperparameter configurations (Supplementary Table 3) were selected based on the mean AUC obtained from this fixed 5-fold setting. To further assess robustness and reduce the effect of random variation, we repeated the entire process under 12 distinct random seeds, yielding a total of 60 runs for each model. The final reported performance metrics represent the mean with standard deviation across all folds and seeds.

Evaluation of model interpretability

To interpret the predictions of our models, we utilized SHAP values³⁸. SHAP values provide a unified measure of variable importance by assigning each variable an importance value for a particular prediction. This method helps in understanding the contribution of each clinical variable to the model’s decision-making process. We computed SHAP values for all variables used in the models to determine their impacts on predicting stable CHD and ACS. By visualizing these SHAP values, we identified which variables had the most significant influence on the model’s predictions. To assess whether seemingly non-significant univariate biomarkers contribute through non-linear patterns or interactions, we computed PDPs and two-way interaction PDPs using the fitted models.

Quantitative assessment

The performance of the proposed model was evaluated by assessing the following metrics, including sensitivity, precision, F1 score, accuracy, and AUC of the ROC.

Sensitivity, precision, F1 score, and accuracy are all related to true-positive (TP), true-negative (TN), false-positive (FP), and false-negative (FN), and are computed as follows,

$\text{Sensitivity}=\frac{\text{TP}}{\text{TP}+\text{FN}}$

$\text{Precision}=\frac{\text{TP}}{\text{TP}+\text{FP}}$

$\text{F}1=\frac{2\times \text{Precision}\times \text{Sensitivity}}{\text{Precision}+\text{Sensitivity}}$

$\mathrm{Accuracy}=\frac{\text{TP}+\text{TN}}{\text{TP}+\mathrm{TN}+\text{FP}+\text{FN}}.$

It is worth noting that when the negative group is relatively large, the model tends to show high accuracy. The ROC curve plots sensitivity against 1-specificity, where specificity$=\frac{\text{TP}}{\text{TP}+\text{FN}}$, and AUC measures the area under the ROC curve.

Computational environment

All experiments were conducted on a high-performance server equipped with eight NVIDIA RTX 4090 GPUs (24 GB memory each), an Intel Xeon Gold 6226R CPU (32 cores), and 768 GB DDR4 RAM, running Ubuntu 20.04 LTS. Models were implemented in PyTorch 2.4.1 with CUDA 12.1 support. Training the T2G-Former model for a single prediction window (6, 12, or 24 months) with a fixed seed and 5-fold cross-validation required approximately 90 min on one GPU. Training across all 12 random seeds and three time windows required around 18 GPU hours in total.

Statistical analysis and sensitivity analysis

Data preprocessing, model training, and statistical analyses were conducted in a reproducible manner using Python packages such as pandas, NumPy, SciPy, and scikit-learn. Descriptive statistics were computed for all clinical and demographic variables to summarize the data, including measures of central tendency and dispersion.

For robustness, we conducted sensitivity analyses, including label-permutation tests to characterize the empirical null distribution of AUC (train-label permutation with retraining, and test-label permutation without retraining) and alternative missing-data strategies (MICE vs k-NN) with full pipeline refitting.

Author contributions

H.M., H.B., J.Y., R.W., and H.X. contributed to the study design. Q.C., Z.M., S.T., and Y.Z. assisted with data curation and assisted with the statistical and machine learning methodology. H.M., H.B., R.W., and H.X. were responsible for conceptualization, formal analysis, investigation, software development, and writing the manuscript. R.W., J.W., and H.X. acted as senior supervisors for all aspects of the project, overseeing management and manuscript revision. All authors read, edited, and approved the final manuscript.

Data availability

The data are not publicly available due to privacy and policy restrictions. For the model, the online platform at https://chd.ai4tomorrow.online will be under continuous development as the authors refine the model and add additional features.

Code availability

All code used in this study is available at https://github.com/HaozhongMa/Ai4CHD.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s44325-025-00094-2.