Deep learning for cardiovascular disease: a comprehensive review of detection and risk forecasting

Abstract

The biggest health threat to the global population is cardiovascular disease (CVD), which afflicts almost one-third of the global population and causes considerable monetary and social losses. Risk analysis should be performed in a timely and appropriate manner to enhance clinical practice and preventive interventions. The emergence of advanced data modalities, such as wearable sensors, medical imaging, electronic health records (EHRs), and genomic platforms, has led to a paradigm shift in the holistic assessment of CVD risk through multimodal data integration. This systematic review is a methodological analysis of recent multimodal input deep-learning algorithms that enhance the early detection of CVD and risk-specific evaluation, following PRISMA 2020 guidelines across 69 studies published 20,122,025 based on 2,847 initial database records. We characterized the wide range of available data streams: longitudinal physiological measurements (ECG, HRV, and BP), echocardiogram data, cardiac MRI and CT, lab/demographic data, behavioral/environmental data, and genomic/proteomic data. Mid-level, early, late, and attention-based fusion methods are described in the context of deep neural networks, such as CNNs, RNNs, BiGRU with attention, and hybrid CNN-LSTM networks. Comparative studies showed dramatic improvements in predictive accuracy (often over 98%), strength to missing or noisy modalities, and access to real-time, individualized recommendations. The best-performing DEEP-CARDIO BiGRU-Attention model had 99.9 percent accuracy on Framingham and Statlog benchmarks. A systematic review of 28 studies by Grad-CAM and SHAP confirmed the dominance of each in imaging and structured-data tasks, respectively (Rahman et al., 2024). The federated explainable FL-LSTM model achieved 99% AUC across three ECG databanks with complete privacy protection. We end with a systematic reproducibility, federated learning, equitable AI, and regulatory translation roadmap.

1 Introduction

Cardiovascular diseases (CVDs) are a group of multiple disorders of the heart and blood vessels, such as coronary artery disease, stroke, heart failure, arrhythmias, and peripheral artery disease, which is the leading cause of death in the world, with an estimated 17.9 million deaths annually, approximating 32 percent of all deaths globally (Holt et al., 2023; Hathaway et al., 2021; Barbieri et al., 2020). In high-income countries, the burden of CVD has decreased with improved healthcare, whereas in low-to-middle income nations, there is a growing risk burden due to metabolic syndrome, urbanization, sedentary lifestyles, and pollution (Wang et al., 2021; Ordikhani et al., 2022) (Figure 1; Table 1).

Figure 1

CVDs do not only have a human cost in terms of premature mortality—30 percent before the age of 70—but also have a significant economic cost in terms of direct medical expenses, lost productivity, and social dependency. The socio-demographic change, rising levels of obesity, sedentary lifestyles, poor diets, pollution, and tobacco use have further increased the number of at-risk populations in all age groups and on all continents, making CVD an urgent global public health concern (Bhagawati et al., 2023; Addissouky et al., 2024). The traditional risk models, namely the Framingham Risk Score and the ASCVD calculator, are based on the assumption of a linear relationship and lack temporal dynamics, limiting generalizability, especially in women, ethnic minorities, and individuals with multimorbidity (Yun et al., 2022; Wang et al., 2021; Ordikhani et al., 2022). Wearable sensors, EHR platforms, omics technologies, and IoT infrastructure are proliferating, revolutionizing CVD prediction by generating multimodal data streams that were previously unavailable to deep neural architectures (Sornalakshmi et al., 2024; Ahmad et al., 2021). The multimodal data landscape includes: (1) Clinical data—laboratory values, diagnoses, comorbidities, medications, and family history; (2) Physiological indicators—longitudinal ECG, HRV, blood pressure, and data from wearable monitors; (3) Medical imaging—echocardiograms, cardiac MRI, CT angiography, and vascular data from mammograms; (4) Genomics/omics—polygenic risk scores, transcriptomics, and methylation signatures; (5) Behavioral/environmental—lifestyle, activity, diet, and exposures sampled using smartphones and apps (Figure 2).

Figure 2

A case in point is systems like DEEP-CARDIO, which streamline the use of deep neural architectures for real-time ECG, blood pressure, pulse, and glucose, classifying conditions and providing interventions in real time. Hybrid neural systems combining CNN and LSTM have achieved predictive accuracies over 98%, compared to traditional systems. Decision trees and SVMs are best applied to structured tabular data, whereas ANNs are used to process mixed clinical data, CNNs to medical images, and RNNs (including LSTM and GRU) on the sequential physiological data (Gupta and Singh, 2023; Zhang et al., 2021; Krishnan et al., 2021; Reshan et al., 2023; Almulihi et al., 2022; Vincent et al., 2022; Yashudas et al., 2024).

1.1 Literature search strategy and PRISMA flow

This review is based on the PRISMA 2020 guidelines. Searches in PubMed, Scopus, IEEE Xplore, Web of Science, and Google Scholar (2015–2025) used Boolean combinations of: deep learning, CVD, multimodal data, ECG, medical imaging, risk prediction, fusion, EHR, wearable sensors, and explainability (Stahlschmidt et al., 2022; Gao et al., 2020). Inclusion criteria: (1) peer-reviewed articles with original DL/ML CVD models; (2) at least one of the following clinical/physiological/imaging/genomic modality; (3) quantitative measures of performance reported. Exclusion criteria: (1) No original model; (2) non-CVD outcomes; (3) lack of methodological detail; (4) non-English; (5) pre-2015 publications; (6) duplicate datasets without new methodology. The first search has yielded 2,847 records in the database and 218 additional sources. After eliminating 312 duplicates, 2,753 records were screened, with 2,341 being eliminated at the title/abstract stage. Out of 412 full-text articles evaluated, 343 were filtered out. Final synthesis: 69 studies, with 59% (n = 41) published 2024–2025 (Singh et al., 2024; Rahman et al., 2024) (Figure 3).

Figure 3

A contemporaneous systematic review of 65 ML studies (Banerjee et al., 2025) confirms that despite high benchmark accuracy, a significant translational gap persists—driven by dataset overreliance, lack of external validation, and black-box opacity—underscoring the necessity of the XAI and multimodal integration focus of the present review.

2 Multimodal data sources in cardiovascular risk stratification

The paradigm shift in predicting CVD is based on integrating five complementary data streams that encode different types of physiological information that cannot be captured adequately by any single data stream (Stahlschmidt et al., 2022; Gao et al., 2020; Zhang et al., 2023; Li et al., 2024). EHR data provide consistent, stable demographic bases; wearable signals capture real-time temporal changes; medical imaging provides spatial morphological details; genomics marks lifetime risk factors; and behavioral/environmental data capture lifetime and exposure determinants (Vincent et al., 2022; Yashudas et al., 2024; Holt et al., 2023). At the center of this paradigm is the compilation of very large, high-resolution databases such as UCI Heart, Framingham, MIMIC-III, PhysioNet MIT-BIH, and Cardiology DICOM, which capture various aspects of cardiovascular risk, including demographic and laboratory data, continuous physiological recordings, and advanced imaging (Figure 4; Tables 2, 3).

Figure 4

2.1 Clinical and structured data

A basic backbone of CVD risk modeling is provided by EHR and structured data (diagnoses, lab values, vital signs, medication use, procedural history, and family history) (Reshan et al., 2023; Bhagawati et al., 2023). The field is supported by reliable publicly available datasets such as UCI Heart Disease (303 records, 14 + features), Framingham Study (4,238 + records), and MIMIC-III (60,000 + ICU admissions) (Table 4).

2.2 Physiological signals and wearable data

Real-time temporal variations, which are not captured by clinical data alone, are provided by physiological signals, namely ECG (at-rest ambulatory), HRV, continuous blood pressure, glucose monitoring, and SpO2 (Vincent et al., 2022; Holt et al., 2023; Krittanawong et al., 2019; Haq et al., 2020). State-of-the-art neural networks, such as RNN, LSTM, GRU, BiLSTM, and BiGRU with attention, have been confirmed on large datasets and have demonstrated exceptional performance in detecting subtle arrhythmias and impending CVD events (Table 5).

The BiGRU-Attention method of the DEEP-CARDIO system was a 99.9% accurate multiclass CVD detector that used simultaneous biosensor data, a significant improvement over classical machine learning or single sensor methods (Krittanawong et al., 2019). The ESMO-optimized EAWO-DNN, by incorporating ESMO optimization, increased network lifetime and diagnostic sensitivity by guaranteeing efficient power clustering in IoT systems (An et al., 2019) (Table 6).

2.3 Medical imaging and radiomics

The imaging-based cardiovascular biomarkers, echocardiography, cardiac magnetic resonance (CMR), coronary computed tomography angiography (CCTA), and mammograms of breast arterial calcification (BAC) allow the knowledge of the spatial, structural, and functional parameters of the cardiovascular system (Gupta and Singh, 2023; Holt et al., 2023; Zhang et al., 2023; Yun et al., 2022). These imaging modalities can give direct, quantifiable measurements of myocardial phenotypes, including chamber size, ejection fraction, vascular wall thickness, plaque burden, and calcification patterns.

Integration of genomics, proteomics, behavioral, and environmental data enables the identification of rare and complex causes, lifetime risk markers, inflammation markers, and social determinants of CVD (Krishnan et al., 2021; Wong and Tse, 2021; Yun et al., 2022). Major repositories such as the UK Biobank and dbGaP support this line of research.

2.4 Genomics, proteomics, and multi-omics

Integration of genomics, proteomics, behavioral, and environmental data enables the identification of rare and complex causes, lifetime risk markers, inflammatory markers, and social determinants of CVD (Krishnan et al., 2021; Wong and Tse, 2021; Yun et al., 2022). Major repositories such as the UK Biobank and dbGaP support this direction of research (Figure 5) (Tables 7, 8).

Figure 5

The EAWO-DNN model shows that energy-sensitive architectures based on ESMO optimization can achieve accuracy beyond 98% and can ensure the life of an IoT network, making continuous real-time monitoring of CVDs clinically feasible (Sornalakshmi et al., 2024). The DEEP-CARDIO system builds upon this by combining streams of wearable biosensor data with EHR in a BiGRU-Attention pipeline and achieves 99.9% accuracy on proven benchmarks (Yashudas et al., 2024).

2.5 Preprocessing, augmentation, and multi-center data harmonization

Robust preprocessing and standardized augmentation are prerequisites for reproducible multimodal CVD modeling. Across the 69 reviewed studies, the following pipelines were consistently applied.

Structured/EHR Data: KNN imputation (k = 5, most common) or median imputation for skewed distributions was used to deal with missing values. Continuous variables (age, BP, cholesterol) were normalized to Z-scores, whereas binary/categorical variables were coded in a one-hot manner. Class imbalance is a widely used concept in datasets such as UCI Heart (54:46) and Framingham (average ratio of events 85:15), as well as in the training of neural networks. Class imbalance is widely observed in datasets such as UCI Heart (54:46) and Framingham (average ratio of events 85:15), as well as in the training of neural networks.

ECG/Physiological Signals: Raw ECG signals were bandpass filtered (0.540 Hz) to remove baseline wander and high-frequency noise, and R-peak detection was done using the Pan-Tompkins algorithm. The signals were grouped into fixed-length windows (usually 5–30 s) with half the window length. The augmentation strategies were: random amplitude scaling (±10%), Gaussian noise addition (SNR 2030 dB), time-warping, and lead dropout simulation to enhance model robustness to real-world signal degradation (Yashudas et al., 2024; Gao et al., 2020).

Medical Imaging: Images obtained by cardiac MRI and echocardiography were resized to a standardized size (224×224 or 112×112 pixels, depending on the architecture). Normalization of intensity was performed per scan, and min-max scaling was used. Standard augmentation was: random horizontal flipping, rotation (± 15 degrees), random cropping, brightness/contrast jitter, and elastic deformation of MRI volumes. In the case of CCTA, windowing (Hounsfield unit) was used [before CNN input (Zhang et al., 2023; Wehbe et al., 2023)] to window Hounsfield unit windowing (Hounsfield unit windowing).

Multi-Center Harmonization: Cross-institutional studies [MIMIC-III, UK Biobank, multi-site federated learning studies (Alasmari et al., 2025; Otoum et al., 2024)] harmonized through: (1) ComBat batch correction of imaging features, (2) feature-level z-score normalization per site, and then fused, (3) federated averaging (FedAvg) to prevent inter-institutional data leakage. Such studies without any protocols for explicit harmonization [Statlog, Cardiology DICOM]. This represents a limitation in reproducibility observed during our quality assessment.

Annotation Standards: The quality of annotations varied significantly across datasets. Multi-reader adjudication, reporting inter-rater reliability (Cohen, 0.80), was used on gold-standard datasets (MIT-BIH, EchoNet Dynamic). Administrative datasets (MIMIC-III) were based on ICD coding that is known to have miscoding rates of 5–15%. Reviewers discovered that 34 of 69 (49%) of those included studies did not report inter-annotator agreement measures, a crucial gap in clinical translation.

3 Deep learning fusion architectures for multimodal CVD risk prediction

It has been demonstrated that the introduction of integrated deep neural networks for multimodal CVD data has established a new gold standard in prediction analytics. New-generation models do not merely incorporate clinical and structured health records, but also include imaging, physiological sensor responses, and molecular profiles to develop a dynamic perspective of disease stratification that is beyond the reach of traditional single-modality models (Stahlschmidt et al., 2022; Gao et al., 2020) (Figure 6).

Figure 6

The underlying rationale is that modalities are statistically complementary: EHR data have stable baselines, wearable signals have temporal dynamics, imaging data have spatial morphology, and genomics data have lifetime risk markers. CNNs are used to analyze images; RNNs (LSTM, GRU) are used to analyze sequential data, whereas decision trees and SVMs are used to analyze structured, tabular data (Figure 7; Table 9).

Figure 7

A deep learning fusion architecture involves working at multiple integration levels. In the first type of fusion (called early fusion), raw feature vectors of all modalities are combined at the input layer; in the second type of fusion (called mid-level fusion), a modality-specific feature-vector is processed independently by separate subnetworks, and then combined with them at a hidden layer; in the third type of fusion (called late fusion), independent model outputs are combined through ensembling (Figure 8; Table 10).

Figure 8

3.1 Comprehensive architecture comparison

Table 11 lists the state-of-the-art models that have been produced by these fusion methods, including the DEEP-CARDIO model with BiGRU-Attention fusion, hybrid CNN-LSTM models, many of which commonly achieve prediction accuracies of 98–99.9%—a noticeable improvement over prior statistical methods.

3.2 Attention mechanisms in CVD prediction

The BiGRU-Attention architecture performs better than hybrids based on CVD risk signals, since these are not stationary; a glucose spike may dominate prediction in one time window, and HRV anomalies in another (Yashudas et al., 2024; An et al., 2019). Attention mechanisms dynamically reweight modalities with respect to each patient state and cannot be done with a static fusion strategy (Figure 9; Table 12).

Figure 9

3.3 Fusion strategy trade-off analysis

The strongest strategy in high-missingness settings is late fusion, since each unimodal model makes an independent prediction that can be combined even when some unimodal models receive no input. Attention fusion clearly increases interpretability through attribution of context-dri (Table 13).

Complementary to these results, synergistic fusion modeling that integrates layers of XAI directly into multimodal pipelines spanning ECG, EHR, and clinical modalities has demonstrated that fusion accuracy and clinical interpretability need not be traded off, achieving about 97.8% accuracy and providing feature-level attribution across these modalities (Figure 10).

Figure 10

4 Validation, benchmarking, and clinical translation

Multimodal CVD AI models should be robustly validated with cross-validation using k-folds, bootstrapping, multi-center external holdouts, and calibration measures such as the Hosmer–Lemeshow test, Brier Score, and Harrell’s C-index. The excellence of multimodal architectures is exhibited in all performance aspects (Table 14).

The proposed DEEP-CARDIO model, when using the BiGRU Attention Network, achieved 99.9% accuracy on wearable + EHR datasets. Hybrid CNN-LSTM networks using IoT-optimized deep neural networks showed consistent high accuracy of more than 98% on large-scale datasets such as Framingham and Statlog, far better standards than legacy statistical and single-modality machine learning models (Oh and Shim, 2024; Wehbe et al., 2023).

So-called dataset-specific preprocessing pipelines, augmentation strategies, and the annotation standards of all major datasets reviewed are summarized in Section 2.5 and Table 15, and provide a reproducible reference frame to all researchers who may wish to replicate or extend the reviewed models.

4.1 Preprocessing and feature engineering

Overall, the preprocessing and feature-engineering techniques used in the reviewed studies are summarized in Table 16, along with their effect on the model performance. For reproductible multimodal CVD modelling, there is a need for standardized pipelines for imputation, normalization and augmentation to minimize inter-study variability and to enable direct comparison of model architectures. These strategies range from structured EHR data, ECG signals, medical imaging to multi-omics inputs, highlighting the variety of modalities that are used in the 69 studies that have been reviewed.

4.2 Architecture selection guidelines by clinical use case

Table 17 shows architecture-selection guidelines correlated to clinical use cases and their corresponding recommendation for the model family, why, and what performance to expect for each deployment scenario. There are many other constraints that dictate the choice of model family beyond performance-maximization, including interpretability requirements, real-time latency, hardware size constraints, and data modality availability, as a few examples. These guidelines are a synthesis of patterns that have been observed on the reviewed literature, for practitioners, who need a practical framework for decisions when deploying multimodal CVD AI in a clinical context.

4.3 Performance comparison across methodological generations

Table 18 compares accuracy and AUC ranges across successive methodological generations, from traditional statistical models to multimodal deep learning architectures. This generational comparison reveals a clear performance staircase: traditional statistical models plateau around 72–87% accuracy, classical ML approaches reach 88–94%, single-modality deep learning achieves 93–97%, and multimodal fusion architectures consistently attain 97–99.9% accuracy on benchmark datasets. Understanding this progression contextualises the specific performance gains attributable to multimodal integration and motivates the architectural design choices documented in the preceding sections.

4.4 Explainability, interpretability, and clinical trust

The multimodal CVD models require a multi-layered interpretability plan to address both imaging and structured data modalities. The systematic XAI review of 28 studies (Rahman et al., 2024) established that Grad-CAM leads in imaging tasks, whereas SHAP leads in structured-data tasks. Bojarczuk et al. (2024) showed that FL-LSTM with SHAP+LIME achieved 99% AUC across three ECG datasets without compromising patient privacy, thereby demonstrating that interpretability and federated privacy can be simultaneously achieved (Figures 11, 12; Table 19).

Figure 11

Figure 12

The XAI methods are further validated across domains and demonstrated to be highly accurate and sensitive in detecting lung cancer, with a commonly used hybrid CNNXGBoost model, DeepXplainer, achieving 97.43% accuracy and 98.71% sensitivity in lung cancer diagnosis, which is the same model with local and global explainability layers that can be replicated and applied to cardiovascular imaging classification tasks with similar black-box trust barriers. In addition to cardiac imaging, cross-domain XAI validation in oncological imaging confirms the generalizability of such methods - with a BiLSTM-CNN + SHAP framework to detect breast cancer, it was demonstrated that composite deep learning with built-in explainability overcomes clinician black-box distrust, which directly applies to CVD imaging AI deployment, where the same barriers of clinician black-box distrust exist.

4.4.1 Grad-CAM—gradient-weighted class activation mapping

Grad-CAM is a method to generate spatial heatmaps by computing the gradient of the class score with respect to the final convolutional feature maps, producing a coarse localization map highlighting regions of the image that most influence the model to make a prediction (Wehbe et al., 2023; Amal et al., 2024; Rahman et al., 2024). Grad-CAM activations in CVD imaging tasks are over myocardial areas of fibrosis, abnormal wall motion, or calcification - which gives cardiologists spatially grounded visual explanations that are consistent with anatomical landmarks they already understand clinically. Quantitative Performance: In cardiac MRI tasks involving segmentation of spatial regions of interest, and echocardiography segmentation tasks, Grad-CAM achieved mean intersection-over-union (IoU) of 0.71 with expert-annotated regions of interest in cardiac MRI tasks, and 0.68 IoU in echocardiography segmentation tasks. Grad-CAM localization achieved similar results for CCTA plaque detection in 78% of true-positive cases (Barbieri et al., 2020; Cocianu et al., 2023). Qualitative Output: Grad-CAM heatmaps of a high-risk cardiac MRI case typically highlight highly active regions of the left ventricular free wall and interventricular septum—areas clinically associated with hypertrophic cardiomyopathy—provide a directly interpretable explanation that can be validated by a cardiologist using standard diagnostic criteria. Limit: Grad-CAM is context-dependent (CNN-only) and produces coarse maps rather than pixel-accurate ones, and is incapable of explaining predictions based on structured/tabular inputs, limiting its applicability in multimodal pipelines where non-imaging data is the driver of prediction.

4.4.2 SHAP—SHapley additive exPlanations

SHAP assigns each input feature a contribution value based on the cooperative game theory Shapley values, to ensure fair additive attribution that satisfies consistency, local accuracy, and missingness axioms (Dharmarathne et al., 2024; Saha et al., 2023; Dehghani et al., 2025). SHAP can be directly applied to clinical risk communication because in CVD structured-data models, it produces both global feature-importance rankings (across the population) and local patient-level explanations (of individual predictions). Quantitative Performance: SHAP feature rankings in the FL-LSTM + SHAP model (Bojarczuk et al., 2024) indicated that serum creatinine, ejection fraction, and age were the top-ranked predictors in all three ECG datasets - in line with known clinical risk factors—with a 99% AUC. The stability analysis of the SHAP features (bootstrap resampling, n = 1,000) confirmed the consistency of the feature ranks with a Spearman correlation of 61. Qualitative Output: A SHAP waterfall plot would look like: age (+0.23), diabetes status (+0.19), troponin level (+0.31), and HDL cholesterol level (−0.14) are the most significant contributors to a high-risk prediction—a cardiologist would follow a direct mapping to clinical logic that would allow transparent shared decision-making. Limitations: SHAP computation is expensive for high-dimensional models [O(2ᴺ) features in exact form], necessitating approximations (KernelSHAP, TreeSHAP) that introduce estimation variance. SHAP additionally presupposes feature independence that is disregarded in correlated clinical data (e.g., BP and age).

4.4.3 LIME—local interpretable model-agnostic explanations

LIME is a locally faithful linear surrogate model that is fitted to any single prediction by perturbing the input and observing the changes in the output, which results in sparse feature attribution explanations that are valid in the local neighborhood of the instance (Dharmarathne et al., 2024; Saha et al., 2023; Bojarczuk et al., 2024). LIME is model-agnostic and applicable to image, text, and tabular CVD data, making it uniquely suited to multimodal explanation pipelines. Quantitative Performance: In paired SHAP+LIME studies (Bojarczuk et al., 2024), convergent validity was confirmed by the top 3 features, identified by both LIME and SHAP across the ECG dataset. Nonetheless, LIME demonstrated greater variance (20.18 vs. SHAP 20.09 in feature ranking) across repeated executions on the same inputs, indicative of its stochastic perturbation sampling. Qualitative Output: In a misclassified example (false negative- high-risk patient predicted to be a low-risk patient), LIME showed that the model over-weighted a normal resting ECG reading and under-weighted the high-risk patient (elevated NT-proBNP biomarker)—an error that can be clinically interpreted as the model missing a biochemical signal absent in the training distribution. Limitations: LIME explanations can only be considered locally valid - they may contradict each other across similar patients and cannot be thought of as global model behavior. The perturbation kernel and neighborhood size are important factors that influence the stability of the outputs.

4.4.4 Quantitative comparison of XAI methods

Table 20 shows quantitative performance comparison of XAI methods in CVD applications.

4.4.5 Clinical case studies—successes and errors

To ground XAI analysis in clinical reality, the following representative case studies synthesize patterns observed across the 28 XAI-evaluated studies.

4.4.5.1 Case study 1—success: Grad-CAM correctly identifies hypertrophic cardiomyopathy (HCM)

Patient Profile: 52-year-old man, without symptoms, who referred to cardiac MRI as a routine procedure. Clinical presentation: normal ECG, mild dyspnea with exercise, family history of sudden cardiac death. Model Prediction: cardiac MRI classifier based on a CNN predicted HIGH RISK (probability = 0.91) for hypertrophic cardiomyopathy. Grad-CAM Explanation: Heatmap demonstrated a strong activation over the basal interventricular septum (septal thickness activation region) and the left ventricular outflow tract - exactly the anatomy landmarks used by cardiologists to diagnose HCM according to ACC/AHA guidelines. Clinical Outcome: The sequential echocardiography showed asymmetric septal hypertrophy (septal wall thickness = 18 mm, >15 mm diagnostic threshold). In a blinded evaluation, the Grad-CAM explanation was evaluated by 3 cardiologists who rated it as either clinically coherent or directly actionable (Wehbe et al., 2023; Amal et al., 2024). XAI Value: The spatial correlation of Grad-CAM activation with known clinical diagnostic criteria (septal thickness) provided a verifiable, trust-building explanation directly supporting clinical decision-making and not necessarily requiring black-box acceptance of the explanation.

4.4.5.2 Case study 2—ERROR: SHAP reveals model bias in female patient cohort

Patient Profile: 61-year-old female, who presents with atypical chest pain, fatigue, and jaw pain. Clinical features: normal troponin, borderline ECG changes, post-menopausal. Multimodal DNN: LOW RISK (probability = 0.12) was predicted by the Multimodal DNN—a false negative. SHAP Explanation: SHAP analysis showed that the model attributed very negative weight to female sex (SHAP value = −0.28) and atypical symptoms (SHAP value = −0.19) which effectively penalizes the patient who presents herself with symptoms that are clinically recognized as the typical CVD presentation pattern among women but underrepresented in male dominated training data (Framingham: 52% female, but CVD event rate 3 times lower in women in training set). Clinical Outcome: The patient was later diagnosed with microvascular coronary artery disease—a disease that is more common in women, and one that has been widely overlooked by models that are trained on primarily male cohorts (Singh et al., 2024; Ayoub et al., 2025). XAI Value: The SHAP error analysis revealed a significant demographic bias—the model had been trained to associate “female sex” with reduced CVD risk, reflecting a training data bias that is not a clinical reality. This case study directly motivated the inclusion of sex-stratified model evaluation and fairness constraints as future research priorities (Singh et al., 2024; Bojarczuk et al., 2024; Ayoub et al., 2025).

4.4.5.3 Case study 3—ERROR: LIME exposes feature leakage in ECG model

Patient Profile: 74-year-old male, with known atrial fibrillation, who was admitted with acute dyspnea. Model Prediction: FL-LSTM model predicted MODERate RISK (probability = 0.54)—inaccurately predicting the actual severity. LIME Explanation: LIME showed the model was giving much weight to HSP timestamp hospital admission (LIME coefficient = +0.31) instead of the clinically significant elevated NT-proBNP and rapid ventricular rate features. Clinical Outcome: Patient needed emergency cardioversion in 6 h—a HIGH RISK outcome that was grossly underestimated by the model. XAI value: The local explanation revealed a spillover error, affecting the overall accuracy of the model (the model was found to exhibit 92% accuracy with this systematic error). As illustrated in this case, XAI can be not only a clinical method of communication but also a model for debugging and quality assurance systems that must be in place prior to clinical implementation (Dharmarathne et al., 2024; Saha et al., 2023; Bojarczuk et al., 2024) (Table 21).

4.5 Result visualization, confusion matrices, ROC/PR analysis, attention heatmaps, and error analysis

Strict presentation of results requires presentation in forms other than scales of accuracy measures. This section presents the synthesis of confusion matrix profiles, ROC and Precision-Recall (PR) curve profiles, attention heatmap profiles, and systematic analysis of errors across the 69 reviewed studies, which offers a clinically based evaluation framework.

4.5.1 Confusion matrix analysis across key models

Confusion matrices disclose the capability of discrimination at the class level that cannot be seen in aggregate measures of accuracy. Table 22 depicts the synthesis of confusion matrix profiles of the best-performing models across the studies reviewed.

Clinical Interpretation: False Negatives (FN) - missed high-risk patients - have the greatest clinical cost in CVD screening. The BiGRU-Attention DEEP-CARDIO model has the lowest FN rate (0.1%) and is thus best suited for population screening, where missed cases are associated with preventable deaths. The increased FN rate of the FL-LSTM model (9.0) reflects the accuracy-privacy trade-off of federated learning, where averaging models across sites yields lower sensitivity than centralized training.

Key Pattern—Class Imbalance Impact: Models trained on UCI Heart Disease (54:46 class ratio) are consistently better balanced in their confusion matrices than models trained on Framingham (85:15 event ratio), where the minority-class confusion decreases without SMOTE or focal loss compensation. Those studies that did not report the use of SMOTE had a mean FN rate 4.2% higher than in studies that reported (Singh et al., 2024; Du et al., 2026).

4.5.2 ROC curve analysis—AUC distribution across model generations

Receiver Operating Characteristic (ROC) curves measure discrimination at all classification thresholds, with AUC summarizing this discrimination in a threshold-independent way in Table 23.

The AUC development through methodological generations demonstrates a clear staircase pattern: the traditional statistical models reach the plateau at AUC 0.720.80; the classical approaches of ML methods reach the plateau at AUC 0.920.96; unimodal DL models reach the plateau at AUC 0.930.98; and multimodal DL models reach the plateau at AUC 0.971.00 (Singh et al., 2024; Rahman et al., 2024). The maximum UA increase occurs at the transition from single-modality to multimodal fusion, supporting the assertion that discrimination gains are driven by modality complementarity, rather than architectural complexity per se. Figure 13a ROC curve comparison across 8 model categories that show AUC staircase of logistic regression (0.76) to BiGRU-Attention (0.997), with shaded 95% confidence intervals per category. The multimodal fusion boundary (AUC > 0.97) is marked by a dashed threshold line, which represents clinically acceptable discrimination for population-level CVD screening.

Figure 13

The threshold should be set above the default 0.5, which is required for clinical deployment. In the case of CVD screening (where FN cost > > FP cost), you can determine the optimum operating point using the Youden Index (J = Sensitivity + Specificity −1). In the case of high-risk alert systems, a threshold of sensitivity of 0.98 is recommended, with a specificity as low as 0.85 (Singh et al., 2024; Rahman et al., 2024).

4.5.3 Precision-Recall (PR) curve analysis—critical for imbalanced CVD datasets

PR curves provide more information than ROC curves in class-imbalanced datasets since they are not sensitive to the large number of true negatives that inflate ROC-AUC in screening populations. Since the event rates in most CVD datasets are 15–54, PR analysis is needed to be honest about them (Table 24).

Narrative behind PR Analysis: BiGRU-Attention, with its dynamic weighting approach, is effective at compensating for the underrepresentation of the minority class on the Framingham dataset (15% event rate) without explicit resampling. Conversely, the FL-LSTM model exhibits the lowest PR performance (AP = 0.905) due to the complexity of multi-class classification across the heterogeneous ECG datasets. Importantly, the models tested on the balanced datasets (EAWO-DNN, UCI-based models) have artificially inflated PR metrics that may not generalize to the real-world clinical populations where event rates are 5–15% (Singh et al., 2024; Du et al., 2026). Figure 13b shown Precision-Recall curves of each of the 6 models overlaid on a single plot overlaid with the random classifier baseline (horizontal line at event rate) shown in reference. The distance between the PR curves of each model and the baseline measure of the actual predictive lift, over and above a chance task, of imbalanced clinical screening tasks.

4.5.4 Attention heatmap analysis

Attention heatmaps are direct visual evidence of which temporal windows, spatial regions, or input features the model finds most discriminative—filling the gap between algorithmic decision-making and clinical reasoning (Table 25).

Heatmap Analysis Narrative: In all 28 XAI-evaluated studies (Rahman et al., 2024), temporal attention heatmaps acquired using BiGRU models consistently identified ST-segment depression windows (lasting 80,120 ms) as the most attentionally active regions of the ECG-based prediction of CVDs—the very diagnostic criteria that cardiologists use to indicate ischemia. The activation patterns in cardiac MRI CNNs showed spatial attention corresponding to clinically relevant anatomical structures and not imaging artifacts (Wehbe et al., 2023; Amal et al., 2024; Rahman et al., 2024). Heatmap Failure Modes: In 5 of 28 studies reviewed (18%), heatmap attention showed high activation over non-pathological regions—that is, artifacts of image acquisition, motion-blur areas from patient motion, and scanner-specific background patterns. These malfunctioning modes are critical patient safety issues, which must be mandatory, clinically verified, and deployed (Rahman et al., 2024).

4.5.5 Error analysis—challenging cases and failure patterns

Systematic error analysis across the reviewed studies reveals five recurring challenging case categories that consistently degrade model performance (Table 26).

Error Analysis Narrative: The most clinically relevant error pattern in all 69 reviewed studies is the systematic underdetection of CVD in female patients and ethnic minorities—a demographic bias due to the composition of training data based on a sample of the population, and not on the inherent architectural constraints of the technology. Models trained on data from only Framingham or UCI show an average FN rate that is 4.8% higher when female patients are used (Singh et al., 2024; Ayoub et al., 2025). This is not an accidental error but a systematic, predictable error that XAI methods (especially SHAP) can detect and alert [as seen in Case Study 2 (Section 4.4.5)]. The second most influential type of error is the modality missingness of early fusion models. In cases of failure in any of the modality streams, wearable sensors disconnection, lack of imaging, or delays in lab results, early fusion architectures demonstrate 8.3% drops in accuracy on average. This drives the architectural suggestion of late fusion or attention-based fusion in clinical deployment settings where the availability of the entire modality cannot be assured (Yashudas et al., 2024; An et al., 2019; Alasmari et al., 2025).

4.5.6 Calibration analysis

In addition to discrimination (AUC) and classification (confusion matrix), clinical deployment also requires calibration, i.e., alignment between predicted probabilities and the actual occurrence rate. A predictive model with 80% CVD risk should be accurate in predicting an 80% risk for the time being (Table 27).

Table 27

Model	Calibration method	Hosmer-Lemeshow p-value	Brier score	Expected Calibration Error (ECE)	Calibration status	References
DEEP-CARDIO BiGRU	Platt scaling post-hoc	p = 0.43 (well calibrated)	0.018	0.022	Well calibrated	Yashudas et al. (2024)
EAWO-DNN	Temperature scaling	p = 0.31	0.021	0.031	Well calibrated	Sornalakshmi et al. (2024)
NSGA-II ensemble	No calibration reported	Not reported	Not reported	Not reported	Unknown	Gupta and Singh (2023)
Physics-guided DL	Physics constraints as implicit prior	p = 0.67	0.031	0.028	Well calibrated	Zhang et al. (2023)
FL-LSTM + SHAP	Federated platt scaling	p = 0.19	0.044	0.051	Moderate calibration	Bojarczuk et al. (2024)

Calibration analysis of top CVD prediction models.

Critical Finding: Of the 69 studies reviewed, only 31 (45%) reported any calibration metrics other than AUC and accuracy. This is a serious gap. A model with 99% accuracy and poor calibration (ECE > 0.10) cannot be safely used to communicate risk to clinicians, as predicted values become meaningless with respect to decision-making by clinicians (Singh et al., 2024; Rahman et al., 2024). The TRIPOD-AI reporting rules require that calibration be reported, and this review recommends that it be a mandatory assessment criterion of all future CVD AI publications.

4.6 Methodological clarity: hyperparameters, training schedules, optimizers, loss functions, and computational environments

To achieve reproducible multimodal CVD deep learning models, there must be full methodological transparency beyond performance measures. This section summarizes the training settings, hyperparameter settings, optimizer strategies, loss functions, and software/hardware environments described in the 69 reviewed studies, which can serve as a reference framework for further researchers.

4.6.1 Hyperparameter configurations

Hyperparameter choices directly govern model capacity, generalization, and convergence behavior. Table 28 consolidates the hyperparameter profiles of the top-performing architectures reviewed. Key parameters documented include learning rate schedules, batch sizes, dropout rates, weight-decay coefficients, and network depth — each of which has a direct impact on whether a model overfits to small benchmark datasets or generalizes to heterogeneous clinical populations. Consistent with reproducibility best practices, this compilation also records whether each study disclosed these settings, revealing notable gaps in methodological transparency.

Key hyperparameter patterns: Among successful models, three consistent hyperparameter choices are identified. First, a dropout rate of 0.3–0.5 is widely used—higher dropout rates (0.4–0.5) in CNN blocks and lower dropout rates (0.2–0.3) in recurrent layers, reflecting the higher propensity of convolutional layers to overfit small medical datasets. Second, the number of hidden units in BiGRU/LSTM follows a funnel architecture (from larger to smaller) with the number of hidden units gradually shrinking. Third, multi-head attention mechanisms with 8 heads are always better than single-head attention in multimodal fusion tasks, thereby achieving cross-modal alignment in multiple subspaces of the representation (Yashudas et al., 2024; Browne et al., 2024).

4.6.2 Training schedules and convergence criteria

Training schedule insights: The most common learning rate strategy that has been reviewed and found effective is ReduceLROnPlateau—reducing the learning rate by a factor of 0.5 when the loss on validation stalls at 10 or more epochs—used by 61% of reviewed studies (Yashudas et al., 2024; Reshan et al., 2023; Browne et al., 2024). The Transformer-based model employs the original warmup schedule (linearly increasing LR 4000 steps then inversely scaling), which is important for achieving stable attention weight initialization. The longest training time (300 epochs) is needed in physics-guided DL as the composite loss (data fidelity + physics constraint) takes more steps to reconcile the competing goals (Zhang et al., 2023). FL-LSTM Training Details: FL-LSTM is trained with 50 local epochs per communication round, distributed across 20 global rounds, 1,000 successful training epochs shared across clients. After each round, FedAvg aggregation occurs, and convergence is determined by stabilization of a global model AUC (< 0.001 improvement over 5 consecutive rounds) (Bojarczuk et al., 2024) (Table 29).

4.6.3 Optimizer choices and rationale

Optimizer analysis: Adam is more likely to optimize CVD DL, which is used in 71% of the reviewed studies, because it has adaptive per-parameter learning rates that can handle the heterogeneous magnitudes of gradients across multimodal inputs (imaging gradients are orders of magnitude larger than tabular feature gradients). RMSprop is one of the choices for models in IoT/wearable applications, where the input distributions vary over time (Sornalakshmi et al., 2024). The LSTM/GRU models require gradient clipping (where the maximum gradient is 1.0–5.0) to ensure that the training process does not experience exploding gradients (Krishnan et al., 2021; Reshan et al., 2023; Yashudas et al., 2024) (Table 30).

4.6.4 Loss functions

Loss function analysis: The most complex imbalance-handling loss function reviewed is Focal Loss [used in FL-LSTM (Bojarczuk et al., 2024)], which down-weights easy negative training examples, and where the γ parameter (set to 2) serves to down-weight hard-to-classify minority-class (positive CVD) training cases. Transformer-CVD label smoothing is used to prevent overconfident predictions, using 0.05/0.95 instead of 0/1, to reduce calibration error and improve generalization (Browne et al., 2024). It is a physics-guided model that uses a composite loss combining two losses: a data-driven BCE and a physics residual loss that penalizes predictions outside the biologically plausible range. This is the most important distinguishing factor that enables the physics-guided approach to yield clinically plausible results (Zhang et al., 2023) (Table 31).

4.6.5 Software frameworks and hardware environments

Software environment observations: PyTorch and TensorFlow are equally represented - 51 vs. 49, respectively, among the reviewed studies, which is a reflection of the plurality of the field of software environments. The research-oriented studies (physics-guided, transformer, federated) are dominated by PyTorch because it supports a dynamic computational graph, enabling custom loss functions and manipulate gradients. Production-oriented IoT studies and clinical deployment studies are dominated by TensorFlow/Keras compatibility with edge hardware (Sornalakshmi et al., 2024; Yashudas et al., 2024). Hardware Gap: 31% of the studies reviewed failed to specify the hardware used to train the model, making it impossible to replicate the training time. The most performance-intensive model, Transformer-CVD on NVIDIA A100 (80GB) in 24 h, is a significant obstacle to the reproducibility of research groups that do not have access to a high-performance computing device (Table 32).

4.6.6 Cross-validation and train/validation/test split protocols

Critical methodological gap—random seed reporting: Only 58% of the reviewed studies reported using fixed random seeds, which is a fundamental requirement of reproducibility. Stochastic components (weight initialization, dropout masks, data shuffling) lack a fixed seed, thereby contributing to non-deterministic behavior across runs and making it impossible to exactly replicate the results. This review suggests seed reporting (training seed, data-split seed, augmentation seed) as a minimum standard of reproducibility of CVD AI publications (Singh et al., 2024; Rahman et al., 2024). Gap of External validation: Only 12 of 69 studies reviewed (17%) conducted external validation using held-out datasets across various institutions. The remaining 83% report only internal cross-validation performance, which is systematically overestimated by 3–8% in AUC, according to meta-analytic estimates (Singh et al., 2024; Banerjee et al., 2025) (Table 33).

4.6.7 Methodological reporting completeness—gap analysis

Gap analysis narrative: The methodological reporting completeness audit indicates that the most important gaps in reproducibility in the reviewed CVD DL literature involve the external validation (17%) and the public code availability (12%). Even though studies report most training parameters, external validation is not established, so the reported benchmark performance figures are likely to be a clinically significant overestimation of actual clinical utility in the real world. This review is the first to recommend external validation, report the random seed, and release code as unnegotiable conditions for future CVD AI publications (Singh et al., 2024; Rahman et al., 2024; Ayoub et al., 2025) (Table 34).

5 Future research directions—accelerating clinical translation of multimodal CVD AI

Although benchmark accuracies exceed 99%, only 3 of 69 reviewed studies (4.3%) report prospective clinical validation, and none report regulatory approval or active clinical deployment. Table 35 summarizes 23 specific research opportunities in five critical directions: multimodal integration, real-time deployment, federated learning, longitudinal studies, and prospective clinical trials. Each has its current gap, proposed approach, expected clinical impact, and implementation timeline.

5.1 Multi-modal integration

Temporal alignment of heterogeneous sampling rates—ECG at 360 Hz, wearables at 1 Hz, labs updated daily, and imaging annually cannot just be concatenated. The highest-priority architectural innovations for real-world clinical applications, where 30–60% of patients have not recorded their data, are temporal alignment networks and missing modality robustness via VAE/diffusion imputation (An et al., 2019; Alasmari et al., 2025).

5.2 Real-time deployment

The critical bottleneck is the accuracy-latency-power trilemma: to be highly accurate, the model must consume substantial GPU resources, which are incompatible with the constraints of wearables. Distillation of knowledge retains 95–97 thousand teacher accuracy at a cost 10 times lower (Sornalakshmi et al., 2024; Yashudas et al., 2024). The regulatory process takes five steps, one after another, where (1) IRB-approved prospective data collection, (2) TRIPOD-AI validation, (3) FDA pre-submission meeting, (4) De Novo/510 (k) application, (5) post-market surveillance; a 4–7 year gap between research and deployment.

5.3 Federated learning

Transatlantic FL networks between hospital systems in North America, Europe, Asia, and Africa would essentially fix the demographic bias issue cited in Section 4.5.5, which is currently barred by GDPR/HIPAA jurisdictional issues. The long-term vision of personalized federated learning via meta-learning (FedPer/MAML) is one in which a continually updated personal CVD risk model is maintained by each patient’s wearable (Otoum et al., 2024; Bojarczuk et al., 2024).

5.4 Longitudinal studies

It is the most important gap in the reviewed literature—all 69 papers are cross-sectional or have a short time window. The longitudinal CVD AI changes the clinical question from whether this patient is at high risk to when this patient is and is not at high risk. It necessitates ongoing wearable monitoring platforms, longitudinal EHR linkage, and time-aware architectures, such as Neural ODEs and temporal transformers (Yashudas et al., 2024; Browne et al., 2024; Yang et al., 2024). The primary source data is the ongoing repeat-imaging study by the UK Biobank.

5.5 Prospective clinical trials

None of the 69 reviewed studies was a prospective RCT, the largest obstacle to clinical adoption. The most urgently needed are three trial designs (1) alert fatigue trial that will test whether AI CVD alerts are more effective in improving physician response than dangerous over-alerts; (2) equity-stratified trial powered to detect superior performance across sex, ethnicity and socioeconomic strata; and (3) federated multi-site RCT that will simultaneously generate clinical evidence and advance the FL methodology (Singh et al., 2024; Ayoub et al., 2025) (Table 36).

Thangaraj et al. (2024) in the European Heart Journal, demonstrated that digital twin technology, a combination of in silico replication of patients with generative AI and multimodal data streams, enables personalized CVD risk simulation and evaluation of a virtual treatment scenario, a paradigm shift from the reactive cardiology paradigm.

6 Conclusion and future suggestions

Multimodal deep learning models, particularly hybrid ensemble architectures and BiGRU-Attention frameworks, have set a new benchmark for CVD risk stratification in both IoT-enabled and clinical settings, consistently achieving over 98% accuracy, precision, and recall. Innovations such as ESMO energy-aware clustering, NSGA-II feature selection, and CNN-GRU-LSTM hybrid architectures effectively address challenges related to data heterogeneity, computational efficiency, and personalized care, thereby enabling continuous monitoring, earlier intervention, scalable precision medicine, and more patient-centered healthcare delivery. Looking ahead, the field must prioritize external validation in diverse, prospective multicenter cohorts to ensure robust generalizability, especially for underrepresented populations; implement federated learning frameworks with differential privacy to enable secure multi-institutional deployment; embed physics-guided modeling and attention-based interpretability directly into model architectures to enhance transparency and clinician trust; and adopt TRIPOD-AI and CONSORT-AI reporting standards as essential requirements for publication and regulatory approval. Additional priorities include developing continuous learning systems that adapt to incoming patient data without catastrophic forgetting, integrating multi-omics data using graph neural networks and multimodal transformers to uncover hidden causal pathways, and advancing digital twin frameworks that merge physiological knowledge with data-driven intelligence for auditable, patient-specific forecasting and virtual clinical trials. Collectively, evidence from 69 studies published between 2012 and 2025, including DEEP-CARDIO BiGRU-Attention (99.9% accuracy), EAWO-DNN for IoT applications (98.9% accuracy, 0.99 AUC), and FL-LSTM with explainable AI (99% AUC across three ECG datasets), demonstrates the rapid evolution of this field. Notably, 59% of these studies were published in 2024–2025, underscoring the accelerating trajectory toward precision cardiovascular medicine.

References

• 1

  AddissoukyT. A.HarbH. A.MohamedS. H.El-SayedA. S.SamyM. (2024). Recent developments in the diagnosis, treatment, and management of cardiovascular diseases through artificial intelligence and other innovative approaches. PRO5, 29–40. doi: 10.46439/biomedres.5.041,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2

  AhmadM.AlfayadM.AftabS.KhanM. A.FatimaA.ShoaibB.et al. (2021). Data and machine learning fusion architecture for cardiovascular disease prediction. Comput. Mater. Contin.69:13. doi: 10.32604/cmc.2021.019013

  • CrossRef
  • Google Scholar

• 3

  AhujaS. K.ShrimankarD. D.DurgeA. R.Design of an Iterative Method for deep multimodal feature fusion in heart disease diagnostics utilizing explainable AI, in Proceeding 1st International Conference Explainable AI for Neural and Symbolic Methods (EXPLAINS 2024), (2024), pp. 87–95.

  • Google Scholar

• 4

  AlasmariS.AlGhamdiR.TejaniG. G.Kumar SharmaS.MousaviradS. J. (2025). Federated learning-based multimodal approach for early cardiac disease detection. Front. Cardiovasc. Med.16:1563185. doi: 10.3389/fphys.2025.1563185,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5

  AlmulihiA. H.AlmulihiA.SalehH.HussienA. M.MostafaS.El-SappaghS.et al. (2022). Ensemble learning based on hybrid deep learning model for heart disease early prediction. Diagnostics12:215. doi: 10.3390/diagnostics12123215,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 6

  AlzahraniA.RazaM. A.AsgharM. Z. (2025). Demystifying diagnosis: An efficient deep learning technique with explainable AI to improve breast Cancer detection. PeerJ Comput. Sci.11:e2806. doi: 10.7717/peerj-cs.2806,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7

  AmalS.SawyerD.KönikA. (2024). Artificial intelligence and multimodal medical imaging data fusion for improving cardiovascular disease care. Front. Radiol.4:1412404. doi: 10.3389/fradi.2024.1412404,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8

  AnY.HuangN.ChenX.WuF. X.WangJ. (2019). High-risk prediction of cardiovascular diseases via attention-based deep neural networks. IEEE/ACM Trans. Comput. Biol. Bioinform.18, 1093–1105. doi: 10.1109/TCBB.2019.2935059,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 9

  AyoubC.CamusY.RoussetJ.DercleD.RaymondD.BosqJ.et al. (2025). Multimodal fusion artificial intelligence model to predict myocarditis and adverse cardiovascular events in cancer patients starting ICI therapy. JACC Advances4:101435. doi: 10.1016/j.jacadv.2024.101435,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 10

  BanerjeeT.ChhabraP.KumarM.KumarA.AbhishekK.ShahM. A. (2025). Pyramidal attention-based T network for brain tumor classification: a comprehensive analysis of transfer learning approaches for clinically reliable and reliable AI hybrid approaches. Sci. Rep.15:28669. doi: 10.1038/s41598-025-11574-x,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 11

  BarbieriS.MehtaS.WuB.BharatC.PoppeK.JormL.et al. (2020). Predicting cardiovascular risk from national administrative databases using a combined survival analysis and deep learning approach. Int. J. Epidemiol.51, 931–944. doi: 10.1093/ije/dyab258,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 12

  BhagawatiM.PaulS.AgarwalS.ProtogeronA.SfikakisP.KitasG.et al. (2023). Cardiovascular disease/stroke risk stratification in deep learning framework: a review. Cardiovasc. Diagn. Therapy12, 557–598. doi: 10.21037/cdt-22-438,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 13

  BojarczukV.SpinosaL. N.de CarvalhoA. C.DietrichJ. L. G. (2024). Hybrid explainable deep learning for heart disease prognosis. Appl. Soft Comput.141.

  • Google Scholar

• 14

  BrowneS.et al. (2024). Transformer-based attention models for cardiovascular risk. Sci. Rep.14.

  • Google Scholar

• 15

  CocianuC.UscatuC.KofidisK.MuraruS.VăduvaA. (2023). Classical, evolutionary, and deep learning approaches of automated heart disease prediction: a case study. Electronics12:663. doi: 10.3390/electronics12071663

  • CrossRef
  • Google Scholar

• 16

  DehghaniS.et al. (2025). Explainable AI in cardiovascular health: multimodal analysis and prediction. IEEE Rev. Biomed. Eng.18, 110–131.

  • Google Scholar

• 17

  DharmarathneG.BogahawaththaM.RathnayakeU.MeddageD. P. P. (2024). Integrating explainable machine learning and user-centric design for cardiovascular disease prediction. Patterns23:200428. doi: 10.1016/j.iswa.2024.200428,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 18

  DuY.HuangK.YueM.ChenC.DengX.WangN. (2026). CASNet: curvature-aware cardiac MRI segmentation with multi-scale and attention-driven encoding for enhanced risk-oriented structural analysis. Front. Med.12:1688872. doi: 10.3389/fmed.2025.1688872,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 19

  El-SofanyH.El-SayedA.TawfikA. H. (2024). A proposed technique for predicting heart disease using feature selection. Sci. Rep.14:23277. doi: 10.1038/s41598-024-74656-2,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20

  Evangelin SoniaS. V.JuliusA.MadusudhanaR. (2023). A multi-modal integrated deep neural network for the prediction of cardiovascular disease. Autom. Remote. Control.

  • Google Scholar

• 21

  FathimaA. J.FaslaM. M. N. (2024). A comprehensive review on heart disease prognostication using different artificial intelligence algorithms. Comput. Methods Biomech. Biomed. Engin.27, 1357–1374. doi: 10.1080/10255842.2024.2319706,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 22

  GaoJ.LiP.ChenZ.ZhangJ. (2020). A survey on deep learning for multimodal data fusion. Neural Comput.32, 829–864. doi: 10.1162/neco_a_01273,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 23

  GuptaA.SinghA. (2023). EDL-NSGA-II: ensemble deep learning framework with NSGA-II feature selection for heart disease prediction. Expert. Syst.40:254. doi: 10.1111/exsy.13254

  • CrossRef
  • Google Scholar

• 24

  HaqI.LiangH.ZengK.WangT.UddinI.LinJ.et al. (2026). Deep learning advancements for cardiovascular diseases (CVDs) diagnosis: imaging modalities, challenges, and future perspectives. Biomed. Signal Process. Control119:109899. doi: 10.1016/j.bspc.2026.109899

  • CrossRef
  • Google Scholar

• 25

  HaqI.et al. (2020). Artificial intelligence in personalized cardiovascular medicine and cardiovascular imaging. Cardiovasc. Diagn. Ther.11, 911–923. doi: 10.21037/cdt.2020.03.09,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 26

  HathawayQ. A.YanamalaN.BudoffM. J.SenguptaP. P.ZebI. (2021). Deep neural survival networks for CVD risk prediction: MESA. Comput. Biol. Med.139:104983. doi: 10.1016/j.compbiomed.2021.104983,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 27

  HoltA.BatinicaB.LiangJ.KerrA.CrengleS.HudsonB.et al. (2023). Development and validation of cardiovascular risk prediction equations in 76,000 people with known cardiovascular disease. Eur. J. Prev. Cardiol.31:314. doi: 10.1093/eurjpc/zwad314

  • CrossRef
  • Google Scholar

• 28

  HoniD. G.OkekeochaS.EdehT. T.JacksonA.EzenwakaC. U. (2024). A one-dimensional convolutional neural network-based method for cardiovascular disease prediction. Inform. Med. Unlocked49:101535. doi: 10.1016/j.imu.2024.101535

  • CrossRef
  • Google Scholar

• 29

  JaltotageB.ChandraratneA.LuciniS.MaharZ. U.RiederM. J.NorgroveL.et al. (2024). Use of artificial intelligence including multimodal systems in cardiovascular disease management. Canad. J. Cardiol. Open40, 1804–1812. doi: 10.1016/j.cjca.2024.07.014,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  KheraR.AlexanderK. P.CutlerM. J.DandekarS. P.BrosiusE. N.FeldmanH. I. (2024). Transforming cardiovascular care with artificial intelligence: from discovery to practice. J. Am. Coll. Cardiol.84, 97–114. doi: 10.1016/j.jacc.2024.05.003,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 31

  KrishnaR.GuptaM.SharmaD.GoyalS. (2023). Ensemble deep learning for multimodal heart disease detection. IEEE Access.

  • Google Scholar

• 32

  KrishnanS.MagalingamP.IbrahimR. (2021). Hybrid deep learning model using recurrent neural network and gated recurrent unit for heart disease prediction. Int. J. Electr. Comput. Eng.11, 5467–5476. doi: 10.11591/ijece.v11i6.pp5467-5476

  • CrossRef
  • Google Scholar

• 33

  KrittanawongC.JohnsonK. W.RosensonR. S.WangZ.AydarM.BaberU.et al. (2019). Deep learning for cardiovascular medicine: a practical primer. Eur. Heart J.40, 2058–2073. doi: 10.1093/eurheartj/ehz056,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 34

  LiY.LiH.MeiF.XiongQ. (2024). A review of deep learning-based information fusion for medical multimodal classification tasks. Comput. Biol. Med.177:108635. doi: 10.1016/j.compbiomed.2024.108635,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 35

  LiY.LiK.WangB.LiuX. (2025). Heart failure prognosis risk assessment model based on multimodal data fusion. Egypt. Inform. J.26.

  • Google Scholar

• 36

  LiX.LiuX.DengX.FanY. (2022). Interplay between artificial intelligence and biomechanics modeling in the cardiovascular disease prediction. Biomedicine10:157. doi: 10.3390/biomedicines10092157,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 37

  MederB.VölzkeC.MaasL. R.MüllerS. E.ThormannT. T.FelzenF. T.et al. (2025). Artificial intelligence to improve cardiovascular population health: state-of-the-art and future directions. Front. Cardiovasc. Med.46, 1907–1916. doi: 10.1093/eurheartj/ehaf125,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 38

  MulaniA.et al. (2025). ML-powered internet of medical things structure for heart disease prediction. J. Pharmacol. Pharmacother.16, 38–45. doi: 10.1177/0976500X241281490

  • CrossRef
  • Google Scholar

• 39

  NarayanY.SinghD. P.BanerjeeT.KourP.RaneK.CA. D. D.et al. (2026). A comparative evaluation of deep learning architectures for prostate cancer segmentation: introducing TrionixNet with N-Core multi-attention mechanism. Arch. Comput. Methods Eng.33, 3707–3746. doi: 10.1007/s11831-025-10411-8

  • CrossRef
  • Google Scholar

• 40

  OgunpolaA.BakareS.OgunpolaO. (2024). Machine learning-based predictive models for detection of heart disease. Diagnostics14:144. doi: 10.3390/diagnostics14020144,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 41

  OhS.ShimJ. Y. (2024). Development and validation of a deep learning–based cardiovascular disease risk prediction model for long-term breast cancer survivors. J. Clin. Oncol.42:12023. doi: 10.1200/JCO.2024.42.16_suppl.12023

  • CrossRef
  • Google Scholar

• 42

  OrdikhaniM.AbadehM. S.PruggerC.HassannejadR.MohammadifardN.SarrafzadeganN. (2022). An evolutionary machine learning algorithm for cardiovascular disease risk prediction. PLoS One17:723. doi: 10.1371/journal.pone.0271723,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43

  OtoumS.et al. (2024). Federated deep learning for smart healthcare IoT: a privacy-preserving CVD screening model. IEEE Internet Things J.

  • Google Scholar

• 44

  RahmanS.et al. (2024). Multimodal data fusion for precision cardiology: current trends and future prospects. BMC Med. Inform. Decis. Mak.

  • Google Scholar

• 45

  ReshanM. S. A.AminS.ZebM. A.SulaimanA.AlshahraniH.ShaikhA. (2023). A robust heart disease prediction system using hybrid deep neural networks. IEEE Access11, 121574–121591. doi: 10.1109/access.2023.3328909

  • CrossRef
  • Google Scholar

• 46

  SadrH.KhalilzadehM. R.RezaeiA.AshrafiN.ZakeriP. (2024). Cardiovascular disease diagnosis: a holistic approach using deep learning and machine learning. Front. Cardiovasc. Med.29:44. doi: 10.1186/s40001-024-02044-7,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  SahaP.et al. (2023). A comprehensive survey on explainable multimodal data fusion for CVD risk stratification. Artif. Intell. Rev.

  • Google Scholar

• 48

  SekarB.DongM.ShiJ.HuX. (2012). Fused hierarchical neural networks for cardiovascular disease diagnosis. IEEE Sensors J.12, 644–650. doi: 10.1109/JSEN.2011.2129506

  • CrossRef
  • Google Scholar

• 49

  SelvarathiC.VaradhaganapathyS. (2023). Deep learning based cardiovascular disease risk factor prediction among type 2 diabetes mellitus patients. Inf. Technol. Control52, 215–227. doi: 10.5755/j01.itc.52.1.32008

  • CrossRef
  • Google Scholar

• 50

  ShaikT.ReddyB. K.KumarA.KumarP. (2024). A survey of multimodal information fusion for smart healthcare. Inf. Fusion102:102040. doi: 10.1016/j.inffus.2023.102040,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 51

  SinghD. P.BanerjeeT.KourP.MalikR.NaiduG. R.CA. D. D.et al. (2026a). A comprehensive study of enhanced computational approaches for breast cancer classification: comparative analysis with existing state of the art methods. Arch. Comput. Methods Eng.33, 3635–3663. doi: 10.1007/s11831-025-10414-5

  • CrossRef
  • Google Scholar

• 52

  SinghD. P.BanerjeeT.MahajanS.Ramesh ChandraK.KumarR.PhaniS.et al. (2026b). A comprehensive study on deep learning models for the detection of diabetic retinopathy using pathological images. Arch. Comput. Methods Eng.33, 503–532. doi: 10.1007/s11831-025-10315-7

  • CrossRef
  • Google Scholar

• 53

  SinghM.GuptaA.VermaD.TiwariN.LodhaJ. (2024). Artificial intelligence for cardiovascular disease risk prediction: review and future directions. J. Cardiovasc. Comput. Tomogr.

  • Google Scholar

• 54

  SornalakshmiM.DevakanthJ. J. M. A.RajalakshmiR.VelmurugadassP. (2024). Energy-aware heart disease prediction using ESMO and optimal deep learning model for healthcare monitoring in IoT. J. Biomol. Struct. Dynamics43, 3542–3556. doi: 10.1080/07391102.2023.2298736,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 55

  StahlschmidtS.et al. (2022). Multimodal deep learning for biomedical data fusion: a review. Brief. Bioinform.23:569. doi: 10.1093/bib/bbab569,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 56

  SumalathaU.PrakashaK.PrabhuS.NayakV. C. (2024). Deep learning applications in ECG analysis and disease detection: an investigation study of recent advances. IEEE Access12, 126258–126284. doi: 10.1109/access.2024.3447096

  • CrossRef
  • Google Scholar

• 57

  ThangarajP. M.BensonS. H.OikonomouE. K.AsselbergsF. W.KheraR. (2024). Cardiovascular care with digital twin technology in the era of generative artificial intelligence. Eur. Heart J.45, 4808–4821. doi: 10.1093/eurheartj/ehae619,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58

  VenkateshC.PrakashS.VenkatP.NarayananR. (2024). An automatic diagnostic model for detection and prognosis of cardiovascular disease using multimodal health record data. Heliyon10:e25574. doi: 10.1016/j.heliyon.2024.e25574,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 59

  VincentP. R.et al. (2022). IoT-cloud-based smart healthcare monitoring system for heart disease prediction via deep learning. Electronics.

  • Google Scholar

• 60

  WangJ.DingH.BidgoliF. A.ZhouB.IribarrenC.MolloiS.et al. (2017). Detecting cardiovascular disease from mammograms with deep learning. IEEE Trans. Med. Imaging36, 1172–1181. doi: 10.1109/TMI.2017.2655486,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 61

  WangY.WuL.ChenZ.NiX.SuM. (2024). Cardiovascular disease prediction model based on patient behavior patterns in deep learning. Front. Psych.15:1418969. doi: 10.3389/fpsyt.2024.1418969,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 62

  WangG.ZhangY.LiS.ZhangJ.JiangD.LiX.et al. (2021). A machine learningbased prediction model for cardiovascular risk in women with preeclampsia. Front. Cardiovasc. Med.8:736491. doi: 10.3389/fcvm.2021.736491,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 63

  WaniN. A.KumarR.BediJ. (2024). DeepXplainer: An interpretable deep learning based approach for lung Cancer detection using explainable artificial intelligence. Comput. Methods Prog. Biomed.243:107879. doi: 10.1016/j.cmpb.2023.107879,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 64

  WehbeR. M.KatsaggelosA. K.HammondK. J.HongH.AhmadF. S.OuyangD.et al. (2023). Deep learning for cardiovascular imaging: a review. JAMA Cardiol.8:3142. doi: 10.1001/jamacardio.2023.3142,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 65

  WongY.TseH. (2021). Circulating biomarkers for cardiovascular disease risk prediction in patients with cardiovascular disease. Front. Cardiovasc. Med.8:713191. doi: 10.3389/fcvm.2021.713191,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 66

  YangH.HuangQ.DuanX.HuangL.WangM. (2024). AI-powered CVD risk modeling from longitudinal multimodal imaging. Front. Radiol.

  • Google Scholar

• 67

  YashudasA.GuptaD.PrashantG. C.DuaA.AlQahtaniD.ReddyA. S. K. (2024). DEEP-CARDIO: recommendation system for cardiovascular disease prediction using IoT network. IEEE Sensors J.24, 14539–14547. doi: 10.1109/jsen.2024.3373429

  • CrossRef
  • Google Scholar

• 68

  YunH.NohN. I.LeeE. Y. (2022). Genetic risk scores used in cardiovascular disease prediction models: a systematic review. Rev. Cardiovasc. Med.23:8. doi: 10.31083/j.rcm2301008,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 69

  ZhangD.ChenY.YeS.CaiW.JiangJ.XuY.et al. (2021). Heart disease prediction based on the embedded feature selection method and deep neural network. J. Healthcare Eng.2021:22. doi: 10.1155/2021/6260022,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 70

  ZhangD.LiuX.XiaJ.GaoZ.ZhangH.de AlbuquerqueV. H. C. (2023). A physics-guided deep learning approach for functional assessment of cardiovascular disease in IoT-based smart health. IEEE Internet Things J.10, 18505–18516. doi: 10.1109/jiot.2023.3240536

  • CrossRef
  • Google Scholar

• 71

  ZhuJ. Y.ZhangL.XiaoS.XuZ.TanQ., Cardiovascular disease detection based on multi-modal data and multi-branch residual networks, in Proceeding ACM International Conference, (2023).

  • Google Scholar

• 72

  ZhuJ.et al. (2025). Dual-scale deep residual network for multimodal cardiovascular disease detection. Comput. Biol. Med.171.

  • Google Scholar