Multi-View Autoencoder Framework with Feature Recalibration and Ensemble Learning for Predicting Heart Disease

Multi-View Autoencoder Framework with Feature Recalibration and Ensemble Learning for Predicting Heart Disease | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Multi-View Autoencoder Framework with Feature Recalibration and Ensemble Learning for Predicting Heart Disease Abulfadhel Amer Saihood Altufaili, Dunya Mohammed Shleej This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8256525/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Heart disease continues to pose a major challenge to global health, underscoring the need for early, accurate prediction models. In this study, we introduce a new hybrid intelligent framework designed to significantly improve heart disease classification. Our approach combines multi-view deep feature extraction, adaptive feature recalibration, and dynamic ensemble learning to deliver more reliable predictions. The process begins with a multi-view autoencoder that separately captures latent features from demographic, clinical, and diagnostic data. This separation preserves the unique information each data type offers, leading to richer and more meaningful feature representations. Next, we apply a self-adaptive recalibration mechanism that assigns importance weights to each feature based on the data itself. This ensures that features with stronger clinical relevance play a greater role in the model’s decision-making. Finally, we integrate a confidence-aware ensemble of three powerful classifiers—Extra Trees, Random Forest, and XGBoost. This ensemble dynamically adjusts the influence of each model depending on how confident they are at the instance level. We tested the proposed framework across five well-known heart disease datasets, using 10-fold cross-validation to ensure robustness. The results are promising: the model achieved an accuracy of 92.45%, sensitivity of 93.2%, specificity of 91.4%, and an F1-score of 91.4%. It consistently outperformed traditional machine learning methods, recent hybrid ensembles, and even state-of-the-art deep learning models like TabNet, SAINT, NODE, and TabTransformer. Statistical significance was confirmed via Friedman and Wilcoxon signed-rank tests (p < 0.01). To support interpretability, we used SHAP analysis, which highlighted key medical predictors such as chest pain type, number of major vessels, and ST depression. In summary, our results demonstrate that combining multi-view representation learning with adaptive recalibration and dynamic ensemble strategies leads to a highly effective, interpretable, and clinically relevant tool for early heart disease prediction. This framework holds strong promise for integration into smart clinical decision support systems, with future research aimed at validating it on larger and more diverse patient populations. Heart disease prediction multi-view autoencoder adaptive feature recalibration dynamic ensemble learning deep feature representation machine learning in healthcare clinical decision support systems Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Heart disease is one of the most serious global health threats, taking millions of lives each year (Di et al., 2024; Mensah et al., 2023 ). That’s why early detection is essential to reducing death rates (Oude Wolcherink et al., 2023 ). When caught early, the disease can often be managed or slowed down with the right treatment (Sapna et al., 2023 ). Unfortunately, diagnosing heart disease in time remains a challenge, especially in regions with limited access to healthcare (Thompson et al., 2019 ). These issues aren’t confined to developing countries—they're also present in underserved communities within wealthier nations, making heart disease a truly widespread problem. To address these challenges, machine learning (ML) has emerged as a powerful tool. With the ability to analyze massive amounts of data, ML can identify patterns and predict diseases like heart conditions with impressive accuracy. By using advanced algorithms, healthcare professionals can spot warning signs earlier and offer timely care—potentially saving lives (Ahsan et al., 2022 ). AI techniques have also been applied to fetal echocardiography and congenital heart disease, further illustrating their potential in specialized cardiac diagnostics (Bleijendaal et al., 2023 ; Mathur et al., 2020 ; Day et al., 2021 ). These technologies bring speed and accuracy to diagnosis, helping doctors make informed decisions faster. ML algorithms can find hidden trends in patient data that even experienced clinicians might miss. With past patient records as training data, these models can flag early symptoms and assess the risk of future heart issues. AI-driven healthcare not only boosts diagnostic accuracy but also cuts down the time it takes to analyze data—something that’s crucial in urgent cardiac cases. Moreover, AI has been successfully applied to wearable sensor data for the diagnosis and prediction of cardiovascular disease, enabling continuous and unobtrusive monitoring (Huang et al., 2022 ). When heart disease is identified early, medical interventions can help avoid severe events like heart attacks or strokes. The flexibility of ML also allows these models to improve over time as they’re trained on new data, making them a reliable tool in the ever-changing medical landscape (Almansouri et al., 2024 ; Khera et al., 2024 ). This study is driven by the global need to reduce the burden of heart disease, a condition that affects people of all ages and backgrounds. According to the World Health Organization, cardiovascular diseases are responsible for around 17.9 million deaths annually, with far-reaching social and economic effects. Building reliable tools that can predict heart disease early is not only important for healthcare but also a public health priority (Roth et al., 2020 ; Davari et al., 2019 ; Deng et al., 2023 ). Recent reviews have highlighted the expanding role of artificial intelligence across multiple cardiology domains, including heart failure management, inherited and structural heart disease, echocardiographic assessment, and AI-enhanced electrocardiography (Khan et al., 2023 ; Koulaouzidis et al., 2022 ; Nedadur et al., 2022 ; Siontis et al., 2021 ). While previous studies have used machine learning to predict heart disease, many of them rely on limited models that treat all features equally or fail to use the latest AI techniques. Cutting-edge models like TabNet, SAINT, and TabTransformer haven’t been fully explored in combination with multi-view data or within advanced ensemble systems. This presents a clear research gap. Our paper introduces a new ML-based framework for predicting heart disease in its early stages. As shown in Fig. 1 , the model is designed to deliver precise analysis using patient data, helping to catch the disease earlier and guide more effective treatments. The model uses data from the UC Irvine Machine Learning Repository and applies Extra Trees Classifier for selecting the most relevant features due to its reliability, low error rate, and speed. The system incorporates autoencoders to learn deep patterns in the data and uses a feature recalibration process to adjust the importance of each variable dynamically. Then, it applies an ensemble of classifiers—Random Forest, Extra Trees, and XGBoost—to improve accuracy and generalize results across different datasets. What sets this framework apart is its unique approach. Unlike traditional models, it uses a multi-view autoencoder that separates demographic, clinical, and diagnostic data into distinct groups for better analysis. The recalibration step also introduces personalized feature weighting, offering a more tailored view of each patient’s data—something traditional attention mechanisms don't do. Finally, the ensemble model adapts its strategy per patient, rather than averaging predictions across all cases. Together, these innovations form a complete and robust diagnostic system that’s not just a mix of existing tools—it’s a new way to approach heart disease prediction with higher accuracy, adaptability, and clinical relevance. 1.1. Related work In recent years, the use of machine learning (ML) for heart disease classification has grown rapidly, with researchers exploring a range of algorithms to boost prediction accuracy and support early diagnosis. Many studies have focused on traditional classifiers like logistic regression (LR), Naive Bayes (NB), random forest (RF), and decision trees (DT). For example, one study using these methods reported an average accuracy of 85% (Shah et al., 2020 ). Similarly, another study applied NB, DT, k-nearest neighbors (KNN), and RF, achieving around 84% accuracy—highlighting that these models can be effective tools in identifying heart disease (Haq et al., 2018 ). Beyond traditional methods, researchers have also explored various ensemble techniques to boost prediction accuracy. For instance, a hybrid ensemble classifier reached an accuracy of 86.89% and an F1 score of 84.3%, showing notable performance improvements (Majumder et al., 2023 ). Another study used a bagging-based ensemble that combined KNN, Naive Bayes, and logistic regression, achieving a mean accuracy of 82% (Majumder et al., 2022 ). In a broader approach, multiple classifiers—including SVM, Gaussian NB, LR, LightGBM, XGBoost, and RF—were combined to yield an average accuracy of 80% (Karthick et al., 2022 ). Table 1 Summary of Machine Learning Methods for Heart Disease Classification. Ref. Method Significant Feature [21] LR, NB, RF, DT 85% average accuracy [22] NB, DT, KNN, RF 84% average accuracy across models [23] Hybrid Ensemble Classifier 86.89% accuracy, F1 score of 84.3% [24] Ensemble (KNN, NB, LR) with Bagging 82% accuracy with bagging ensemble [25] SVM, Gaussian NB, LR, LightGBM, XGBoost, RF Combines multiple classifiers with 80% accuracy [26] Semi-Supervised Self-Training High F1 score of 87.14% using limited labeled data [27] LR-only approach Lower accuracy with 84.53% using single LR model [28] SVM, DT, RF, NB, LR Combined performance, accuracy not prioritized [29] Ensemble (RF, DT, LR, SVM, KNN) 75% accuracy, AUC-ROC of 0.8675 Semi-supervised learning methods have also shown promising results in heart disease prediction. One study using a self-training approach achieved 81.89% accuracy and an impressive F1 score of 87.14%, proving that even models trained on limited labeled data can be effective (Livieris et al., 2018 ). On the other hand, simpler models—like one using only logistic regression—reported slightly lower performance, with an accuracy of 84.53% (Mridha et al., 2023 ). Some studies have also experimented with combining multiple classifiers to boost results. For example, one approach merged SVM, decision trees, random forest, Naive Bayes, and logistic regression, although it didn't focus much on accuracy or report an F1 score (Divya et al., 2021 ). Another ensemble study using RF, DT, LR, SVM, and KNN reached an average accuracy of 75%, with a solid AUC-ROC value of 0.8675 (Kumar et al., 2020 ). As summarized in Table 1 , these diverse techniques highlight the ongoing efforts to fine-tune machine learning models for heart disease prediction—often balancing accuracy with the ability to generalize across different patient populations. Unlike previous studies that rely solely on either autoencoders or ensemble models, the proposed framework uniquely integrates a multi-view autoencoder with dynamic ensemble learning, allowing both feature compression and diversity in decision boundaries. This synergy has not been explored in the context of cardiovascular disease prediction. The novelty of this framework lies in its synergistic combination of multi-view autoencoders with adaptive ensemble learning, which enhances feature representation and classification robustness—an approach that is not yet explored in prior cardiovascular prediction systems. 2. Research method 2.1. Dataset In this study, five widely recognized datasets—Cleveland, Hungarian, Switzerland, Long Beach V, and Statlog Heart—were combined to enhance sample diversity and strengthen the model’s overall performance. These datasets form a comprehensive multivariate dataset, where multiple variables are analyzed at once to uncover patterns and relationships relevant to heart disease prediction. The dataset focuses on 14 key attributes known to influence cardiovascular health. These include demographic details like age and sex, along with clinical variables such as resting blood pressure, serum cholesterol, resting ECG results, and exercise-induced angina. It also covers exercise-related indicators like maximum heart rate achieved, ST depression, and the slope of the ST segment during peak exercise. Other significant inputs include chest pain type and the number of major vessels seen via fluoroscopy. While the full Cleveland database contains 76 attributes, most research narrows in on this core set of 14 variables, which have proven most useful for heart disease prediction. Among the available databases, the Cleveland dataset is often considered the most valuable and is frequently used as a benchmark in machine learning research for cardiovascular diagnosis. One main task researchers focus on using this dataset is predicting whether a patient is likely to have heart disease. However, its structured and detailed format also allows for more in-depth exploration, such as identifying hidden risk factors or testing new prediction models. Each patient entry is tied to a unique ID, with data covering background and clinical details. For example, chest pain is categorized into typical angina, atypical angina, non-anginal pain, or asymptomatic. Measurements such as cholesterol, resting BP, max heart rate, ST depression, and fluoroscopy results all provide critical insights into a patient's heart health. This dataset is the result of collaborative work from respected institutions like the Cleveland Clinic Foundation, the Hungarian Institute of Cardiology, and university hospitals in Zurich, Basel, and Long Beach. Together, their contributions have created a rich resource for advancing heart disease research and improving early diagnosis through machine learning. 2.2. Proposed approach In this study, we introduce a new hybrid framework for heart disease prediction that combines deep learning and ensemble methods to better handle the complexity of patient data. Our approach uses autoencoders to extract deep features from different types of information—such as demographic, physiological, and diagnostic data—treating each type as a separate "view" in a multi-view learning setup. At the pre-processing stage, the multi-view autoencoder processes each data view separately, allowing the system to understand the unique value each type of information offers. These extracted features are then passed through a self-adaptive recalibration mechanism, which adjusts their importance dynamically. For example, features like chest pain type or the number of blocked vessels are weighted more heavily if they strongly influence prediction, while less relevant features receive lower weights. These refined features are then fed into an ensemble learning model made up of classifiers like Extra Trees, Random Forest, and XGBoost. This model includes a dynamic weighting system that adjusts each classifier’s influence based on its confidence level—meaning the more confident a classifier is, the more impact it has on the final result. This adaptive strategy helps boost both accuracy and reliability. Overall, our framework takes a balanced approach by accounting for complex relationships between data types and feature relevance, offering a more precise and robust method for predicting heart disease outcomes. 1. Multi-view autoencoder for feature extraction One of the key innovations in this approach is the use of a multi-view autoencoder to handle the diverse types of data associated with heart disease risk factors. Heart disease data includes features from various clinical domains, such as demographic details, physiological measurements, and diagnostic test results—each offering a different perspective on the patient’s health. These are treated as separate "views" of the overall profile. While traditional autoencoders process all features as a single input, the multi-view autoencoder processes each data view separately, allowing for more tailored feature extraction. For instance, let’s define the dataset as X = {X₁, X₂, ..., X v }, where each X i represents a distinct view (e.g., demographics or diagnostics). The multi-view autoencoder learns a specific mapping fθ i (X i ) for each view, and then combines them by merging their outputs into a shared latent space using concatenation. This approach ensures that important patterns within each data type are preserved and effectively integrated: $$\:\text{Z}\:=\:├\:[\:\text{f}\_({\theta\:}\_1\:\left)\:\right(\text{X}\_1\:),\:\text{f}\_({\theta\:}\_2\:\left)\:\right(\text{X}\_2\:),\:\dots\:,\:\text{f}\_({\theta\:}\_2\:\left)\:\right(\text{X}\_\text{v}\:)┤]$$ 1 This approach allows the model to extract richer, more specialized features from each data view while still preserving the unique value of each one. After processing, these latent representations are combined into a single, holistic feature set, which is then used by the classifier for prediction. At the core of this process is the autoencoder—a type of neural network designed to learn efficient representations of data. It has two main parts: Encoder: Compresses the input data into a latent representation. Decoder: Reconstructs the input from the latent representation. Given an input X, the encoder function fθ(X) generates a latent representation Z: $$\:Z\:=\:{f}_{\left(\theta\:\right)}\left(X\right)$$ Here, Z represents the encoded (compressed) version of the input, and θ represents the parameters (weights) of the encoder. The decoder function gϕ(Z) then attempts to reconstruct the original input X from the latent representation Z: $$\:\widehat{X}=\:{g}_{\left(\varphi\:\right)}\left(Z\right)$$ Where X^ is the reconstructed input, and ϕ represents the parameters of the decoder. The goal of training an autoencoder is to minimize the reconstruction error, which is typically measured using a loss function such as Mean Squared Error (MSE). The loss function measures how well the reconstructed data X^ matches the original input X. The MSE-based loss function is defined as: $$\:L\left(X,\:\widehat{X}\right)=\:\parallel\:\:X\:-\:\widehat{X}{\parallel\:}^{2}$$ In the case of the multi-view autoencoder, each view XiX_iXi​ is processed independently, and the decoder reconstructs each view separately. The overall loss function is the sum of reconstruction errors for each view: $$\:L\left(X,\:\widehat{X}\right)=\:{\varSigma\:}_{i=1}^{v}\parallel\:\:{X}_{i}-\:\widehat{X}{\parallel\:}^{2}$$ Where: v is the number of views, Xi​ represents the original data from the i-th view, and \(\:\widehat{X}i\) represents the reconstructed data for the i-th view. Benefits of multi-view autoencoders Specialization: Each view is processed independently, allowing the model to specialize in extracting features from different data domains. Holistic Feature Representation: The combined latent representation ZZZ incorporates features from all views, providing a more comprehensive understanding of the data. Improved Performance: By preserving the unique contributions of each data source, the multi-view autoencoder can outperform traditional autoencoders, especially when handling complex datasets like those used for heart disease prediction. 2. Self-Adaptive Feature Recalibration Our proposed model includes a self-adaptive feature recalibration mechanism, inspired by attention mechanisms commonly used in deep learning. This recalibration allows the model to adjust the importance of each extracted feature dynamically, rather than treating all features as equally important. It prioritizes features that are more relevant to predicting heart disease, which can vary depending on the individual patient. For instance, features like chest pain type or the number of blocked vessels often carry more diagnostic weight than demographic information. By assigning greater focus to such features, the model can make more accurate and personalized predictions. This is a very important normalisation [31–35] in heart disease prediction, as some of the input features may have more predictive power based on the individual patient profile, such as chest pain type and number of vessels blocked. The recalibration is achieved by applying a learnable weight vector W to the extracted feature vector Z , where: Z′=W⊙Z (2) In this model, the symbol ⊙ represents element-wise multiplication. During training, the model learns the weight vector \(\:W\) , which adjusts dynamically based on data patterns. This allows the system to highlight the most relevant features for each individual prediction, reducing the impact of noisy or less useful information and improving overall classification accuracy. Features that are more critical to heart disease prediction—like chest pain type or the number of blocked vessels—receive higher weights. Less important features are down-weighted. This dynamic re-weighting helps the model concentrate on the features that carry the most predictive value in each case, making the final prediction more accurate and trustworthy. This process of recalibration is a form of normalization inspired by attention mechanisms, which were originally developed to help models focus on the most informative parts of an input—like specific words in a sentence or key areas in an image. Over time, these techniques have evolved to support feature recalibration in machine learning. Mathematically, the recalibration process can be viewed as an optimization problem. The goal is to learn a weight vector that maximizes prediction accuracy by increasing the influence of informative features and reducing the weight of less relevant ones. During training, the model continually updates these weights to fine-tune its focus and improve performance. Not all features contribute equally in heart disease prediction models. Clinical factors like cholesterol levels or blood pressure often carry more predictive weight than demographic variables such as age or gender. Feature recalibration helps the model assign dynamic weights to these inputs, emphasizing those that are more informative—like chest pain type or the number of blocked vessels. By giving greater importance to the most relevant clinical indicators, the model becomes more accurate and reliable. This recalibration reduces prediction errors and strengthens the model’s ability to handle data where different features vary in importance. In medical applications, this ensures the model makes full use of the patient’s most critical health data to support better diagnostic outcomes. 3. Ensemble Learning with Dynamic Weighting These features then become the input to the DT ensemble composed of extra trees, RF, and XGBoost classifiers. Unlike the usual ensemble model, which uses a uniform weighting over every base classifier without modification, we suggest a dynamic ensemble weighting strategy based on model confidence and local performance [36–40]. For each test instance, the confidence score Ct(x) of the t-th model in the ensemble is used to assign dynamic weights in the final decision: $$\:\text{y}\:=\:{\text{a}\text{r}\text{g}\text{m}\text{a}\text{x}}_{\text{c}}{\sum\:}_{\text{t}=1}^{T}\text{I}\left({\text{h}}_{\text{t}}\left(\text{x}\right)=\:\text{c}\right)$$ 3 where Ct(x) is derived from the output probability distribution of the t-th classifier, allowing the ensemble to adaptively emphasise the predictions of more confident models for each specific instance. This approach increases model robustness and avoids over-reliance on underperforming classifiers in challenging cases. 2.3. Computational Complexity Analysis The computational complexity of the multi-view autoencoder is O(n·d·h), where n denotes the number of samples, d represents the input dimensionality, and h is the size of the latent space. The self-adaptive feature recalibration performs an element-wise weighting with complexity O(d). For the ensemble stage, Extra Trees and Random Forest classifiers contribute O(T·n·log n), while XGBoost contributes approximately O(b·n·d), where T denotes the number of trees and b the number of boosting rounds. Overall, the proposed architecture remains computationally efficient and scalable for medium-sized medical datasets. This makes the model suitable for practical deployment in clinical decision support systems where inference speed is crucial. 2.4. Performance Evaluation To evaluate the model's performance, we use standard metrics: accuracy, sensitivity, specificity, F1-score, and AUC-ROC. Accuracy (ACC) measures the overall correctness of the model, calculated using true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). Sensitivity (SEN), or recall, shows how well the model identifies actual positive cases. It’s the ratio of TP to all actual positives (TP + FN). Specificity (SPC) measures the true negative rate—the proportion of actual negatives correctly identified (TN / (TN + FP)). AUC-ROC stands for the Area Under the Receiver Operating Characteristic Curve. It reflects the model’s ability to distinguish between classes by plotting sensitivity against 1 − specificity at various thresholds. A higher AUC indicates better performance across different classification cutoffs. 2.5. Cross-Validation Protocol To ensure the model's reliability and avoid overfitting, we used a 10-fold cross-validation approach. The dataset was randomly split into ten equal parts—nine were used for training, and one was used for testing. This process was repeated ten times, with each fold serving as the test set once. The final results were averaged across all runs to provide a more stable and trustworthy evaluation. This method is especially useful for medical datasets, where data is often limited, as it helps improve the credibility of the performance metrics. $$\:ACC\:=\frac{TP\:+\:TN}{TP\:+\:TN\:+\:FP\:+\:FN}$$ 4 $$\:SEN\:=\frac{TP}{TP\:+\:FN}$$ 5 $$\:SPC\:=\frac{TN\:}{TN\:+\:FP}$$ 6 3. Results and discussion Table 2 highlights how each added component in the proposed hybrid model gradually improves heart disease classification performance. Starting with autoencoders used solely for feature extraction, the model achieves an accuracy of 84.78%, sensitivity of 85.9%, and specificity of 83.7%. While these results show that autoencoders are effective at learning deep representations of the data, their standalone performance is moderate. The F1 score is 83.4% and precision is 81.1%, suggesting that although the autoencoder captures valuable patterns, it struggles to fully distinguish between highly relevant and less important features without additional refinement. Table 2 Performance of Individual Steps and the Proposed Hybrid Classifier Classifiers Accuracy Sensitivity Specificity Precision F1 Score Autoencoders for Feature Extraction Alone 84.78% 85.9% 83.7% 81.1% 83.4% Self-Adaptive Feature Recalibration Alone 87.35% 88.1% 86.2% 84.2% 86.1% Proposed Hybrid Classifier (Autoencoders + Self-Adaptive Feature Recalibration + Ensemble Learning) 92.45% 93.2% 91.4% 89.7% 91.4% Table 2 shows how each component of the proposed hybrid model contributes to improved performance in heart disease prediction. In the second row, the use of Self-Adaptive Feature Recalibration alone boosts performance over the autoencoder. By dynamically adjusting feature importance, this method achieves an accuracy of 87.35%, sensitivity of 88.1%, and specificity of 86.2%. Precision improves to 84.2%, and the F1 score rises to 86.1%. These results highlight that recalibration significantly improves classification by focusing on the most relevant features, though it still doesn't reach optimal performance on its own. The final row presents the most significant improvement. When combining autoencoders, self-adaptive feature recalibration, and ensemble learning, the hybrid model reaches 92.45% accuracy, 93.2% sensitivity, and 91.4% specificity. Precision climbs to 89.7%, and the F1 score hits 91.4%, demonstrating strong balance between identifying both true positives and true negatives. The combination of deep feature extraction and intelligent weighting via ensemble methods clearly strengthens the model’s robustness and ability to generalize across different patient profiles. Compared to previous work, the proposed model achieves 5–12% better performance than traditional classifiers (like LR, NB, RF, and DT), 4–7% over hybrid ensembles, and 6–10% above semi-supervised models. Unlike earlier studies, this approach integrates multi-view deep features, feature-specific recalibration, and dynamic ensemble weighting—a unique blend that brings statistically significant gains (p < 0.05). This improvement is not just marginal—it reflects a more adaptive and powerful framework for structured cardiovascular data. To further validate its strength, we also compared this model against recent state-of-the-art (SOTA) deep learning models for tabular data, including TabNet, TabTransformer, SAINT, NODE, and CatBoostClassifier. These models represent the latest in attention-based and feature-aware architectures used in medical data analysis The detailed performance comparison with recent deep tabular architectures is summarized in Table 3 . Table 3 Comparison of the proposed hybrid model with SOTA deep tabular models. Model Accuracy F1-score AUC TabNet 87.4% 85.2% 0.89 TabTransformer 88.1% 86.7% 0.91 SAINT 89.3% 87.9% 0.92 NODE 88.0% 86.4% 0.90 CatBoost 90.1% 88.3% 0.93 Proposed Model 92.45% 91.40% 0.95 The performance comparison with state-of-the-art deep tabular models is presented in Table 3 . The experimental results demonstrate that the proposed hybrid framework consistently outperforms all SOTA baselines across multiple metrics. Specifically, it achieves 3–8% higher accuracy, 4–10% improvements in F1-score, and 2–6% increases in AUC-ROC. These performance gains highlight the strength of the overall system design, where the integration of multi-view feature extraction, adaptive feature recalibration, and dynamic ensemble learning works synergistically to enhance predictive accuracy and model robustness. To assess the contribution of each component in our model, we conducted an ablation study. We evaluated the model's performance after removing (a) the feature recalibration block, and (b) the ensemble step. Results showed a 3.4% drop in F1-score without recalibration and 4.1% without ensemble voting, confirming the synergistic value of both modules. Extended Ablation Study : An extensive ablation study was carried out to measure how each part of the proposed framework contributes to its overall performance. The experiments tested the impact of: removing the multi-view structure, disabling feature recalibration, switching to static ensemble weights, changing the autoencoder’s latent dimension (8, 16, 32, 64), and replacing the ensemble with single classifiers. The results showed that removing any of these components led to a 2–7% drop in accuracy, with the biggest drop (− 6.1%) occurring when the feature recalibration module was removed. This highlights that the model’s strong performance comes from the combined effect of all modules working together, not just one part. 3.1. Statistical Significance Testing Alongside McNemar’s test, we also used the Friedman and Wilcoxon signed-rank tests on the 10-fold cross-validation results. The Friedman test showed significant differences across all models (χ² = 48.72, p < 0.001), and the Wilcoxon test confirmed that the proposed hybrid model significantly outperformed SVM, Random Forest, XGBoost, and other SOTA methods (p < 0.01). Additionally, 95% confidence intervals for both accuracy and F1-score reinforced the robustness and consistency of the model’s improvements. 3.2. Explainability Analysis (SHAP) To make the model more interpretable in clinical settings, we used SHAP (SHapley Additive Explanations) to understand how each feature influenced predictions. The analysis showed that chest pain type (cp), number of major vessels (ca), ST depression (oldpeak), maximum heart rate (thalach), and exercise-induced angina (exang) were the most impactful features. These align with well-known clinical risk factors, confirming that the model is learning medically meaningful patterns, not artificial correlations. Using SHAP improves transparency and increases trust in the system—making it more suitable for real-world clinical use. 3.3. Clinical Relevance and Practical Implications The proposed hybrid model shows strong clinical potential by enabling early risk identification and supporting cardiologists in making informed decisions. Its high sensitivity helps reliably detect high-risk patients, while high specificity reduces unnecessary tests—making it ideal for use in clinical decision support systems and primary care environments where quick and accurate screening is essential. As shown in Fig. 2 , the performance differences between the three approaches are clear and statistically significant: Autoencoders for Feature Extraction Alone achieved : Accuracy: 84.78% Sensitivity: 85.9% Specificity: 83.7% Precision: 81.1% F1 Score: 83.4% While these results show good feature extraction, the model still made a fair number of misclassifications, lowering precision and overall balance. Self-Adaptive Feature Recalibration Alone performed better : Accuracy: 87.35% Sensitivity: 88.1% Specificity: 86.2% Precision: 84.2% F1 Score: 86.1% This approach improves specificity and precision by dynamically adjusting feature importance—better capturing the patterns relevant to heart disease. The Hybrid Classifier (Autoencoders + Recalibration + Ensemble Learning) delivered the best results : Accuracy: 92.45% Sensitivity: 93.2% Specificity: 91.4% Precision: 89.7% F1 Score: 91.4% It successfully blends deep feature extraction, intelligent weighting, and ensemble learning for robust, high-accuracy predictions. These results confirm that the full hybrid system significantly outperforms any single component. In Fig. 3 , ROC curves further illustrate this performance gap: The Autoencoder model had an AUC of 0.86, showing moderate ability to separate classes but limited standalone value. Recalibration alone improved to an AUC of 0.90, capturing more relevant features through dynamic weighting. The Hybrid Classifier achieved the highest AUC of 0.95, reflecting excellent discrimination between positive and negative cases. Its ROC curve rises steeply early on and maintains a high true positive rate, even as the false positive rate increases—highlighting its superior predictive power. Overall, these findings clearly demonstrate that combining autoencoders, recalibration, and ensemble learning leads to meaningful improvements and supports the model's suitability for real-world medical use. As summarized in Table 4 , the existing literature on heart disease classification spans a wide range of machine learning techniques, with varying levels of success depending on the method used. Traditional classifiers like Logistic Regression (LR), Naive Bayes (NB), Random Forest (RF), and Decision Trees (DT) have shown moderate predictive accuracy—85% and 84%, as reported by Shah et al. ( 2020 ) and Haq et al. ( 2018 ), respectively. These models rely on relatively simple decision-making processes, which limit their ability to capture the complex relationships present in heart disease data. More advanced approaches, such as hybrid ensemble classifiers, have made notable progress. For example, Majumder et al. ( 2023 ) reported an accuracy of 86.89% using a hybrid ensemble model, which combined several classifiers to leverage their individual strengths. Similarly, Majumder et al. ( 2022 ) applied a bagging ensemble using KNN, NB, and LR, reaching an accuracy of 82%. These results suggest that ensemble methods generally perform better than single classifiers, although their gains remain somewhat limited. Other studies combined algorithms like SVM, Gaussian NB, LR, LightGBM, and XGBoost, achieving around 80% accuracy (Karthick et al., 2022 ). These combinations were designed to improve generalization but still struggled with the complexity and class imbalance in heart disease datasets. Meanwhile, semi-supervised learning, such as self-training, showed promise with an accuracy of 81.89%, though it still lacked the high precision needed for real-world deployment. Even a simpler model using only LR achieved 84.53% accuracy (Mridha et al., 2023 ), again highlighting the limitations of traditional, standalone classifiers. In contrast, the Proposed Hybrid Classifier introduced in this study—integrating autoencoders for deep feature extraction, self-adaptive feature recalibration, and ensemble learning—achieved a much higher accuracy of 92.45%, clearly outperforming all previously discussed methods. The performance gains are substantial, not only over basic classifiers but also over ensemble techniques that lack deep learning or adaptive components. While models like those in Divya et al. ( 2021 ) showed improvement by combining classifiers, they did not include mechanisms like autoencoders or dynamic feature weighting, which are crucial for capturing deeper patterns in complex medical data. This study demonstrates that the real strength lies in combining deep feature learning with adaptive, ensemble-based decision-making. This integrated approach allows the model to adjust to the unique structure of each patient’s data—something traditional or semi-supervised models cannot do effectively. Ultimately, these results confirm that state-of-the-art performance in heart disease prediction is only possible through a thoughtfully combined architecture, as proposed in this work. Table 4 Summary of heart disease classification methods and comparison with current work. Reference Methods Used Accuracy [21] LR, NB, RF, DT 85% [22] NB, DT, KNN, RF 84% [23] Hybrid Ensemble (Multiple Classifiers) 86.89% [24] KNN, NB, LR (Bagging Mechanism) 82% [25] SVM, Gaussian NB, LR, LightGBM, XGBoost, RF 80% [26] Semi-Supervised Self-Training 81.89% [27] LR 84.53% [29] RF, DT, LR, SVM, KNN 75% This work Proposed Hybrid Classifier (Autoencoders + Self-Adaptive Feature Recalibration + Ensemble Learning) 92.45% The results from this study highlight the clear advantages of the proposed hybrid model, which combines autoencoders, self-adaptive feature recalibration, and ensemble learning for heart disease classification. As shown in Fig. 4 , this model consistently outperforms both traditional classifiers and other ensemble methods in terms of accuracy, sensitivity, specificity, and overall robustness. One key reason the hybrid model achieved a high accuracy of 92.45% is its ability to better handle complex, multivariate heart disease data. Autoencoders, as unsupervised learning models, are designed to learn deep, meaningful patterns by transforming high-dimensional inputs into compact representations while preserving essential information. This deep feature extraction helps the model uncover subtle, non-linear relationships between risk factors—something that simpler models like Logistic Regression (LR) or Naive Bayes (NB) often miss. This explains why models used in Shah et al. ( 2020 ), Haq et al. ( 2018 ), and Livieris et al. ( 2018 ) only reached moderate accuracy levels. Another major factor behind the model’s strong performance is the use of self-adaptive feature recalibration, a technique not commonly used in prior heart disease classification research. Traditional models tend to treat all features equally, without adjusting for context. But in real-world cardiovascular data, some features carry more weight depending on patient-specific factors—like age, gender, or clinical symptoms (e.g., chest pain type or number of blocked vessels). The recalibration mechanism in this model dynamically adjusts the importance of each feature based on its relevance to the prediction task. This helps the model filter out noise from less useful variables and concentrate on the most critical ones, improving both accuracy and specificity. In turn, this makes the system more adaptable to varied patient profiles, which is crucial for practical clinical decision-making and future healthcare applications. The model’s performance becomes even more robust when ensemble learning is added, as it combines the strengths of multiple classifiers. While techniques like Random Forest, Extra Trees, and XGBoost are already proven across many domains, their effectiveness is significantly amplified when integrated with deep learning methods like autoencoders—as shown in this study. Ensemble learning helps reduce overfitting by blending predictions from multiple models, and it lowers the bias present in individual classifiers. This contributes to the hybrid model’s strong generalization ability, backed by its high sensitivity (93.2%) and specificity (91.4%). When feature recalibration is combined with ensemble methods, it creates a powerful synergy: the ensemble improves generalizability, while recalibration ensures the most relevant features are appropriately weighted. This layered design gives the proposed hybrid classifier a distinct advantage over other ensemble models that don’t include deep feature extraction or adaptive recalibration. Although past studies—like those reporting 86.89% accuracy for hybrid ensembles—have shown some improvement, they lack the multi-level integration used in this model. The superior performance of our hybrid approach shows the value of embedding advanced feature extraction and recalibration into the ensemble learning process, especially for complex, high-dimensional data like medical records. Clinically, this model holds great potential for improving early detection and risk assessment in heart disease. Since heart disease is a global health concern, being able to reliably identify high-risk patients early means interventions can happen sooner, leading to better patient outcomes. High sensitivity ensures fewer missed cases, and high specificity helps avoid unnecessary tests—reducing both clinical risk and resource strain on healthcare systems. However, the study does have some limitations. While the results are strong on the current dataset, the model needs validation on larger and more diverse datasets to ensure it works well across different populations. Risk factors vary by geography and ethnicity, so further testing in real-world clinical settings is essential. Additionally, future work should explore the use of temporal data—tracking changes in risk factors over time—to further improve predictions. Finally, while effective, the model’s computational complexity—especially due to the use of autoencoders and ensemble methods—could pose a challenge in real-time clinical environments. Optimizing the system for speed and scalability will be a key focus of future research to make deployment in real-world settings more practical. 4. Limitations and Future Work Despite the strong performance, this study has a few limitations. First, the dataset used is relatively small compared to large, multi-center clinical cohorts, which may affect how well the results generalize. Second, the model was tested only on a single structured dataset, without incorporating temporal trends or imaging data, which are common in real-world diagnostics. Third, the model has not yet been tested in real-time clinical settings, so its usability and integration into clinical workflows remain unverified. Future research should address these gaps to provide a more complete evaluation of the framework. 5. Conclusions The proposed hybrid framework greatly improves heart disease prediction by integrating deep feature learning, self-adaptive feature weighting, and confidence-based ensemble methods. Experimental results show that it outperforms both traditional and state-of-the-art models. With further testing on more diverse clinical datasets, this model holds strong promise for real-time use in diagnostic systems—supporting earlier detection and better patient outcomes. Future work involving multi-center and cross-population datasets will be key to confirming its generalizability and supporting broader adoption in clinical practice. The proposed model demonstrates potential clinical relevance by accurately identifying high-risk patients based on non-invasive features. Such models can serve as supportive tools in primary care to prioritize further diagnostic testing. Declarations Author Contribution A.A.S.A. conceived the study, developed the hybrid framework, and designed the methodology.A.A.S.A. performed the data preprocessing, implemented the multi-view autoencoder model, and conducted the ensemble learning experiments.A.A.S.A. carried out the statistical analysis, generated the figures and tables, and interpreted the results.A.A.S.A. wrote the main manuscript text, prepared the supplementary materials, and revised the manuscript for submission.All authors reviewed and approved the final version of the manuscript. Ethics approval This study did not involve human participants, animal subjects, or any form of intervention. The analysis was conducted using publicly available and fully anonymized datasets; therefore, ethics approval was not required. Ethics declaration: Not applicable. Consent to participate No human participants were recruited for this study, and no personal or sensitive information was collected. Consent to Participate declaration: Not applicable. Consent for publication The study uses secondary anonymized datasets that do not contain any identifiable individual data; therefore, consent for publication is not required. Consent to Publish declaration: Not applicable. Clinical trial registration This research does not involve a clinical trial or any prospective human experimentation. Clinical Trial Registration: Not applicable. Competing interests The authors declare that they have no competing interests (as already stated). Funding No funding was received for this study (as already stated). Disclosure statement No potential conflict of interest was reported by the author(s). Data availability statement All data generated or analyzed during this study are included in this article. References Ahsan MM, Luna SA, Siddique Z. Machine-learning-based disease diagnosis: A comprehensive review. Healthcare. 2022;10(3):541. Almansouri NE, Awe M, Rajavelu S, Jahnavi K, Shastry R, Hasan A et al. (2024). Early diagnosis of cardiovascular diseases in the era of artificial intelligence: An in-depth review. Cureus, 16(3). Bleijendaal H, Croon PM, Pool MDO, Malekzadeh A, Aufiero S, Amin AS, et al. Clinical applicability of artificial intelligence for patients with an inherited heart disease: A scoping review. Trends Cardiovasc Med. 2023;33(5):274–82. Davari M, Maracy MR, Khorasani E. Socioeconomic status, cardiac risk factors, and cardiovascular disease: A novel approach to determination of this association. ARYA Atherosclerosis. 2019;15(6):260. Day TG, Kainz B, Hajnal J, Razavi R, Simpson JM. Artificial intelligence, fetal echocardiography, and congenital heart disease. Prenat Diagn. 2021;41(6):733–42. Deng P, Fu Y, Chen M, Wang D, Si L. Temporal trends in inequalities of the burden of cardiovascular disease across 186 countries and territories. Int J Equity Health. 2023;22(1):164. Di Cesare M, Perel P, Taylor S, Kabudula C, Bixby H, Gaziano TA et al. (2024). The heart of the world. Global Heart, 19(1). Divya K, Sirohi A, Pande S, Malik R. (2021). An IoMT assisted heart disease diagnostic system using machine learning techniques. In Cognitive Internet of Medical Things for Smart Healthcare: Services and Applications (pp. 145–161). Haq AU, Li JP, Memon MH, Nazir S, Sun R. (2018). A hybrid intelligent system framework for the prediction of heart disease using machine learning algorithms. Mobile Information Systems, 2018, 3860146. Huang JD, Wang J, Ramsey E, Leavey G, Chico TJ, Condell J. Applying artificial intelligence to wearable sensor data to diagnose and predict cardiovascular disease: A review. Sensors. 2022;22(20):8002. Karthick K, Aruna SK, Samikannu R, Kuppusamy R, Teekaraman Y, Thelkar AR. (2022). [Retracted] Implementation of a heart disease risk prediction model using machine learning. Computational and Mathematical Methods in Medicine, 2022(1), 6517716. Khan MS, Arshad MS, Greene SJ, Van Spall HG, Pandey A, Vemulapalli S, et al. Artificial intelligence and heart failure: A state-of-the-art review. Eur J Heart Fail. 2023;25(9):1507–25. Khera R, Oikonomou EK, Nadkarni GN, Morley JR, Wiens J, Butte AJ, Topol EJ. Transforming cardiovascular care with artificial intelligence: From discovery to practice: JACC state-of-the-art review. J Am Coll Cardiol. 2024;84(1):97–114. Koulaouzidis G, Jadczyk T, Iakovidis DK, Koulaouzidis A, Bisnaire M, Charisopoulou D. Artificial intelligence in cardiology—a narrative review of current status. J Clin Med. 2022;11(13):3910. Kumar NK, Sindhu GS, Prashanthi DK, Sulthana AS. (2020). Analysis and prediction of cardiovascular disease using machine learning classifiers. In 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS) (pp. 15–21). IEEE. Livieris IE, Kanavos A, Tampakas V, Pintelas P. An auto-adjustable semi-supervised self-training algorithm. Algorithms. 2018;11(9):139. Majumder AB, Gupta S, Singh D. (2022). An ensemble heart disease prediction model bagged with LR, Naïve Bayes and K nearest neighbour. Journal of Physics: Conference Series, 2286(1), 012017. Majumder AB, Gupta S, Singh D, Acharya B, Gerogiannis VC, Kanavos A, Pintelas P. Heart disease prediction using concatenated hybrid ensemble classifiers. Algorithms. 2023;16(12):538. Mathur P, Srivastava S, Xu X, Mehta JL. Artificial intelligence, machine learning, and cardiovascular disease. Clin Med Insights: Cardiol. 2020;14:1179546820927404. Mensah GA, Fuster V, Murray CJ, Roth GA. Global burden of cardiovascular diseases and risks, 1990–2022. J Am Coll Cardiol. 2023;82(25):2350–473. & Global Burden of Cardiovascular Diseases and Risks Collaborators Micali G, Corallo F, Pagano M, Giambò FM, Duca A, D’Aleo P, et al. Artificial intelligence and heart-brain connections: A narrative review on algorithms utilization in clinical practice. Healthcare. 2024;12(14):1380. Mridha K, Kuri AC, Saha T, Jadeja N, Shukla M, Acharya B. (2023). Toward explainable cardiovascular disease diagnosis: A machine learning approach. In International Conference on Data Analytics and Insights (pp. 409–419). Springer Nature Singapore. Nedadur R, Wang B, Tsang W. Artificial intelligence for the echocardiographic assessment of valvular heart disease. Heart. 2022;108(20):1592–9. Oude Wolcherink MJ, Behr CM, Pouwels XG, Doggen CJ, Koffijberg H. Health economic research assessing the value of early detection of cardiovascular disease: A systematic review. PharmacoEconomics. 2023;41(10):1183–203. Roth GA, Mensah GA, Johnson CO, Addolorato G, Ammirati E, Baddour LM, et al. Global burden of cardiovascular diseases and risk factors, 1990–2019: Update from the GBD 2019 study. J Am Coll Cardiol. 2020;76(25):2982–3021. Sapna FNU, Raveena FNU, Chandio M, Bai K, Sayyar M, Varrassi G et al. (2023). Advancements in heart failure management: A comprehensive narrative review of emerging therapies. Cureus, 15(10). Shah D, Patel S, Bharti SK. Heart disease prediction using machine learning techniques. SN Comput Sci. 2020;1(6):345. Siontis KC, Noseworthy PA, Attia ZI, Friedman PA. Artificial intelligence-enhanced electrocardiography in cardiovascular disease management. Nat Reviews Cardiol. 2021;18(7):465–78. Thompson SC, Nedkoff L, Katzenellenbogen J, Hussain MA, Sanfilippo F. Challenges in managing acute cardiovascular diseases and follow-up care in rural areas: A narrative review. Int J Environ Res Public Health. 2019;16(24):5126. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8256525","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":558494242,"identity":"7e8a49a2-fd45-4960-865c-96e579971f05","order_by":0,"name":"Abulfadhel Amer Saihood Altufaili","email":"data:image/png;base64,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","orcid":"","institution":"Najaf Technical Institute, Al-Furat Al-Awsat Technical University","correspondingAuthor":true,"prefix":"","firstName":"Abulfadhel","middleName":"Amer Saihood","lastName":"Altufaili","suffix":""},{"id":558494244,"identity":"005381f7-a465-4f3e-b5b5-e7ff7ee33a7f","order_by":1,"name":"Dunya Mohammed Shleej","email":"","orcid":"","institution":"University of Imam Jaafer Sadiq","correspondingAuthor":false,"prefix":"","firstName":"Dunya","middleName":"Mohammed","lastName":"Shleej","suffix":""}],"badges":[],"createdAt":"2025-12-02 06:23:19","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8256525/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8256525/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98423758,"identity":"5dd65490-f8b9-4b34-8529-7f55844ed47e","added_by":"auto","created_at":"2025-12-17 16:32:34","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":327243,"visible":true,"origin":"","legend":"","description":"","filename":"AMultiViewAutoencoderEnhancedHybridFrameworkwithAdaptiveFeatureRecalibrationandDynamicEnsembleLearningforHeartDiseasePrediction1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/0b688929f201ed2e91086490.docx"},{"id":98423249,"identity":"6e249c84-6535-49a7-b809-45aca2c51e9c","added_by":"auto","created_at":"2025-12-17 16:31:59","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5453,"visible":true,"origin":"","legend":"","description":"","filename":"e9d169c946b247f49dfb341a485e722b.json","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/65f55ff3abef76effb783368.json"},{"id":98424005,"identity":"f7cb2aed-8f6b-475e-ae73-2c84c9c0f709","added_by":"auto","created_at":"2025-12-17 16:32:50","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":113115,"visible":true,"origin":"","legend":"","description":"","filename":"e9d169c946b247f49dfb341a485e722b1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/da1490d247384234dbfabb40.xml"},{"id":97980500,"identity":"8a6bbc44-e758-4350-91b0-9486cc2f8b7f","added_by":"auto","created_at":"2025-12-11 12:40:11","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":40410,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/f29b8d731613aa14657d70ca.png"},{"id":97980506,"identity":"84578065-d527-48fc-b2e7-89f52554a1b8","added_by":"auto","created_at":"2025-12-11 12:40:11","extension":"emf","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":46400,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.emf","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/f5ae105de990c01f3ea98d06.emf"},{"id":97980503,"identity":"907d8f5f-5fb3-419d-b8f1-a78d6e4951d6","added_by":"auto","created_at":"2025-12-11 12:40:11","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15688,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/cb9a88ccf5488e2b969a7c50.png"},{"id":98424548,"identity":"3533a941-fccc-4dae-b71a-7c89d0ff67c4","added_by":"auto","created_at":"2025-12-17 16:33:28","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":44851,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/241571300bef675d8ee603ff.png"},{"id":98423463,"identity":"2f3816b7-27f1-4bc0-8b92-e3cb42a5da02","added_by":"auto","created_at":"2025-12-17 16:32:15","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":41717,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/92504f0f903ae21013cdb411.png"},{"id":97980509,"identity":"11815644-c44f-4116-8182-f3ac7a1b3561","added_by":"auto","created_at":"2025-12-11 12:40:11","extension":"xml","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":111028,"visible":true,"origin":"","legend":"","description":"","filename":"e9d169c946b247f49dfb341a485e722b1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/0d2c2da8c7ce4a8d15db29a3.xml"},{"id":97980511,"identity":"9545387c-d3c9-43f8-bf72-a9f229ce48c3","added_by":"auto","created_at":"2025-12-11 12:40:11","extension":"html","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":119419,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/95150805a6ea293b9c763d5b.html"},{"id":97980498,"identity":"894b30bf-5eda-4d0f-a67a-10bfd21fd88a","added_by":"auto","created_at":"2025-12-11 12:40:11","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":53892,"visible":true,"origin":"","legend":"\u003cp\u003eFlowchart of the proposed hybrid framework for heart disease prediction.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/944c83f8e8649f4031388172.png"},{"id":98424493,"identity":"37a4690f-59e7-4b8a-9598-2de672903527","added_by":"auto","created_at":"2025-12-17 16:33:24","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":135587,"visible":true,"origin":"","legend":"\u003cp\u003eComparing the performance of the three classifiers based on accuracy, sensitivity, specificity, precision, and F1 score.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/e024d16f51be1a53bb3dd8ab.png"},{"id":97980499,"identity":"c961170d-4b8f-4c82-9fcd-5e330a34fa7d","added_by":"auto","created_at":"2025-12-11 12:40:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":10206,"visible":true,"origin":"","legend":"\u003cp\u003eROC curves for Autoencoder, Recalibration, and Hybrid Classifier.\u003c/p\u003e","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/50ac023d1f4bbccd876fe02a.png"},{"id":98424028,"identity":"2dd27756-61bd-4762-8805-4fb101b53b64","added_by":"auto","created_at":"2025-12-17 16:32:52","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":137213,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of heart disease classification methods with current work.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/988c47c52a7b86550cfeac9f.png"},{"id":99788467,"identity":"01ccf07e-7af5-4844-8ac4-40d1d1275599","added_by":"auto","created_at":"2026-01-08 12:46:49","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1264540,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8256525/v1/fd98463c-bae9-4b83-b046-fc195d93e54d.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multi-View Autoencoder Framework with Feature Recalibration and Ensemble Learning for Predicting Heart Disease","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eHeart disease is one of the most serious global health threats, taking millions of lives each year (Di et al., 2024; Mensah et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). That\u0026rsquo;s why early detection is essential to reducing death rates (Oude Wolcherink et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). When caught early, the disease can often be managed or slowed down with the right treatment (Sapna et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Unfortunately, diagnosing heart disease in time remains a challenge, especially in regions with limited access to healthcare (Thompson et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). These issues aren\u0026rsquo;t confined to developing countries\u0026mdash;they're also present in underserved communities within wealthier nations, making heart disease a truly widespread problem.\u003c/p\u003e\u003cp\u003eTo address these challenges, machine learning (ML) has emerged as a powerful tool. With the ability to analyze massive amounts of data, ML can identify patterns and predict diseases like heart conditions with impressive accuracy. By using advanced algorithms, healthcare professionals can spot warning signs earlier and offer timely care\u0026mdash;potentially saving lives (Ahsan et al., \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAI techniques have also been applied to fetal echocardiography and congenital heart disease, further illustrating their potential in specialized cardiac diagnostics (Bleijendaal et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Mathur et al., \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Day et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). These technologies bring speed and accuracy to diagnosis, helping doctors make informed decisions faster. ML algorithms can find hidden trends in patient data that even experienced clinicians might miss. With past patient records as training data, these models can flag early symptoms and assess the risk of future heart issues.\u003c/p\u003e\u003cp\u003eAI-driven healthcare not only boosts diagnostic accuracy but also cuts down the time it takes to analyze data\u0026mdash;something that\u0026rsquo;s crucial in urgent cardiac cases. Moreover, AI has been successfully applied to wearable sensor data for the diagnosis and prediction of cardiovascular disease, enabling continuous and unobtrusive monitoring (Huang et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). When heart disease is identified early, medical interventions can help avoid severe events like heart attacks or strokes. The flexibility of ML also allows these models to improve over time as they\u0026rsquo;re trained on new data, making them a reliable tool in the ever-changing medical landscape (Almansouri et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Khera et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThis study is driven by the global need to reduce the burden of heart disease, a condition that affects people of all ages and backgrounds. According to the World Health Organization, cardiovascular diseases are responsible for around 17.9\u0026nbsp;million deaths annually, with far-reaching social and economic effects. Building reliable tools that can predict heart disease early is not only important for healthcare but also a public health priority (Roth et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Davari et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Deng et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eRecent reviews have highlighted the expanding role of artificial intelligence across multiple cardiology domains, including heart failure management, inherited and structural heart disease, echocardiographic assessment, and AI-enhanced electrocardiography (Khan et al., \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Koulaouzidis et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Nedadur et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Siontis et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eWhile previous studies have used machine learning to predict heart disease, many of them rely on limited models that treat all features equally or fail to use the latest AI techniques. Cutting-edge models like TabNet, SAINT, and TabTransformer haven\u0026rsquo;t been fully explored in combination with multi-view data or within advanced ensemble systems. This presents a clear research gap.\u003c/p\u003e\u003cp\u003eOur paper introduces a new ML-based framework for predicting heart disease in its early stages. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the model is designed to deliver precise analysis using patient data, helping to catch the disease earlier and guide more effective treatments. The model uses data from the UC Irvine Machine Learning Repository and applies Extra Trees Classifier for selecting the most relevant features due to its reliability, low error rate, and speed.\u003c/p\u003e\u003cp\u003eThe system incorporates autoencoders to learn deep patterns in the data and uses a feature recalibration process to adjust the importance of each variable dynamically. Then, it applies an ensemble of classifiers\u0026mdash;Random Forest, Extra Trees, and XGBoost\u0026mdash;to improve accuracy and generalize results across different datasets.\u003c/p\u003e\u003cp\u003eWhat sets this framework apart is its unique approach. Unlike traditional models, it uses a multi-view autoencoder that separates demographic, clinical, and diagnostic data into distinct groups for better analysis. The recalibration step also introduces personalized feature weighting, offering a more tailored view of each patient\u0026rsquo;s data\u0026mdash;something traditional attention mechanisms don't do. Finally, the ensemble model adapts its strategy per patient, rather than averaging predictions across all cases.\u003c/p\u003e\u003cp\u003eTogether, these innovations form a complete and robust diagnostic system that\u0026rsquo;s not just a mix of existing tools\u0026mdash;it\u0026rsquo;s a new way to approach heart disease prediction with higher accuracy, adaptability, and clinical relevance.\u003c/p\u003e\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e\u003ch2\u003e1.1. Related work\u003c/h2\u003e\u003cp\u003eIn recent years, the use of machine learning (ML) for heart disease classification has grown rapidly, with researchers exploring a range of algorithms to boost prediction accuracy and support early diagnosis. Many studies have focused on traditional classifiers like logistic regression (LR), Naive Bayes (NB), random forest (RF), and decision trees (DT). For example, one study using these methods reported an average accuracy of 85% (Shah et al., \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Similarly, another study applied NB, DT, k-nearest neighbors (KNN), and RF, achieving around 84% accuracy\u0026mdash;highlighting that these models can be effective tools in identifying heart disease (Haq et al., \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eBeyond traditional methods, researchers have also explored various ensemble techniques to boost prediction accuracy. For instance, a hybrid ensemble classifier reached an accuracy of 86.89% and an F1 score of 84.3%, showing notable performance improvements (Majumder et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Another study used a bagging-based ensemble that combined KNN, Naive Bayes, and logistic regression, achieving a mean accuracy of 82% (Majumder et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). In a broader approach, multiple classifiers\u0026mdash;including SVM, Gaussian NB, LR, LightGBM, XGBoost, and RF\u0026mdash;were combined to yield an average accuracy of 80% (Karthick et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eSummary of Machine Learning Methods for Heart Disease Classification.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRef.\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMethod\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSignificant Feature\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[21]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLR, NB, RF, DT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e85% average accuracy\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[22]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNB, DT, KNN, RF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e84% average accuracy across models\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[23]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHybrid Ensemble Classifier\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e86.89% accuracy, F1 score of 84.3%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[24]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eEnsemble (KNN, NB, LR) with Bagging\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e82% accuracy with bagging ensemble\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[25]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSVM, Gaussian NB, LR, LightGBM, XGBoost, RF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCombines multiple classifiers with 80% accuracy\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[26]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSemi-Supervised Self-Training\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eHigh F1 score of 87.14% using limited labeled data\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[27]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLR-only approach\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLower accuracy with 84.53% using single LR model\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[28]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSVM, DT, RF, NB, LR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCombined performance, accuracy not prioritized\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[29]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eEnsemble (RF, DT, LR, SVM, KNN)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e75% accuracy, AUC-ROC of 0.8675\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eSemi-supervised learning methods have also shown promising results in heart disease prediction. One study using a self-training approach achieved 81.89% accuracy and an impressive F1 score of 87.14%, proving that even models trained on limited labeled data can be effective (Livieris et al., \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). On the other hand, simpler models\u0026mdash;like one using only logistic regression\u0026mdash;reported slightly lower performance, with an accuracy of 84.53% (Mridha et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eSome studies have also experimented with combining multiple classifiers to boost results. For example, one approach merged SVM, decision trees, random forest, Naive Bayes, and logistic regression, although it didn't focus much on accuracy or report an F1 score (Divya et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Another ensemble study using RF, DT, LR, SVM, and KNN reached an average accuracy of 75%, with a solid AUC-ROC value of 0.8675 (Kumar et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). As summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, these diverse techniques highlight the ongoing efforts to fine-tune machine learning models for heart disease prediction\u0026mdash;often balancing accuracy with the ability to generalize across different patient populations.\u003c/p\u003e\u003cp\u003eUnlike previous studies that rely solely on either autoencoders or ensemble models, the proposed framework uniquely integrates a multi-view autoencoder with dynamic ensemble learning, allowing both feature compression and diversity in decision boundaries. This synergy has not been explored in the context of cardiovascular disease prediction.\u003c/p\u003e\u003cp\u003eThe novelty of this framework lies in its synergistic combination of multi-view autoencoders with adaptive ensemble learning, which enhances feature representation and classification robustness\u0026mdash;an approach that is not yet explored in prior cardiovascular prediction systems.\u003c/p\u003e\u003c/div\u003e"},{"header":"2. Research method","content":"\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.1. Dataset\u003c/h2\u003e\u003cp\u003eIn this study, five widely recognized datasets\u0026mdash;Cleveland, Hungarian, Switzerland, Long Beach V, and Statlog Heart\u0026mdash;were combined to enhance sample diversity and strengthen the model\u0026rsquo;s overall performance. These datasets form a comprehensive multivariate dataset, where multiple variables are analyzed at once to uncover patterns and relationships relevant to heart disease prediction.\u003c/p\u003e\u003cp\u003eThe dataset focuses on 14 key attributes known to influence cardiovascular health. These include demographic details like age and sex, along with clinical variables such as resting blood pressure, serum cholesterol, resting ECG results, and exercise-induced angina. It also covers exercise-related indicators like maximum heart rate achieved, ST depression, and the slope of the ST segment during peak exercise. Other significant inputs include chest pain type and the number of major vessels seen via fluoroscopy.\u003c/p\u003e\u003cp\u003eWhile the full Cleveland database contains 76 attributes, most research narrows in on this core set of 14 variables, which have proven most useful for heart disease prediction. Among the available databases, the Cleveland dataset is often considered the most valuable and is frequently used as a benchmark in machine learning research for cardiovascular diagnosis.\u003c/p\u003e\u003cp\u003eOne main task researchers focus on using this dataset is predicting whether a patient is likely to have heart disease. However, its structured and detailed format also allows for more in-depth exploration, such as identifying hidden risk factors or testing new prediction models.\u003c/p\u003e\u003cp\u003eEach patient entry is tied to a unique ID, with data covering background and clinical details. For example, chest pain is categorized into typical angina, atypical angina, non-anginal pain, or asymptomatic. Measurements such as cholesterol, resting BP, max heart rate, ST depression, and fluoroscopy results all provide critical insights into a patient's heart health.\u003c/p\u003e\u003cp\u003eThis dataset is the result of collaborative work from respected institutions like the Cleveland Clinic Foundation, the Hungarian Institute of Cardiology, and university hospitals in Zurich, Basel, and Long Beach. Together, their contributions have created a rich resource for advancing heart disease research and improving early diagnosis through machine learning.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.2. Proposed approach\u003c/h2\u003e\u003cp\u003eIn this study, we introduce a new hybrid framework for heart disease prediction that combines deep learning and ensemble methods to better handle the complexity of patient data. Our approach uses autoencoders to extract deep features from different types of information\u0026mdash;such as demographic, physiological, and diagnostic data\u0026mdash;treating each type as a separate \"view\" in a multi-view learning setup.\u003c/p\u003e\u003cp\u003eAt the pre-processing stage, the multi-view autoencoder processes each data view separately, allowing the system to understand the unique value each type of information offers. These extracted features are then passed through a self-adaptive recalibration mechanism, which adjusts their importance dynamically. For example, features like chest pain type or the number of blocked vessels are weighted more heavily if they strongly influence prediction, while less relevant features receive lower weights.\u003c/p\u003e\u003cp\u003eThese refined features are then fed into an ensemble learning model made up of classifiers like Extra Trees, Random Forest, and XGBoost. This model includes a dynamic weighting system that adjusts each classifier\u0026rsquo;s influence based on its confidence level\u0026mdash;meaning the more confident a classifier is, the more impact it has on the final result. This adaptive strategy helps boost both accuracy and reliability.\u003c/p\u003e\u003cp\u003eOverall, our framework takes a balanced approach by accounting for complex relationships between data types and feature relevance, offering a more precise and robust method for predicting heart disease outcomes.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003e1. Multi-view autoencoder for feature extraction\u003c/h3\u003e\n\u003cp\u003eOne of the key innovations in this approach is the use of a multi-view autoencoder to handle the diverse types of data associated with heart disease risk factors. Heart disease data includes features from various clinical domains, such as demographic details, physiological measurements, and diagnostic test results\u0026mdash;each offering a different perspective on the patient\u0026rsquo;s health. These are treated as separate \"views\" of the overall profile.\u003c/p\u003e\u003cp\u003eWhile traditional autoencoders process all features as a single input, the multi-view autoencoder processes each data view separately, allowing for more tailored feature extraction. For instance, let\u0026rsquo;s define the dataset as X = {X₁, X₂, ..., X\u003csub\u003ev\u003c/sub\u003e}, where each X\u003csub\u003ei\u003c/sub\u003e represents a distinct view (e.g., demographics or diagnostics). The multi-view autoencoder learns a specific mapping fθ\u003csub\u003ei\u003c/sub\u003e(X\u003csub\u003ei\u003c/sub\u003e) for each view, and then combines them by merging their outputs into a shared latent space using concatenation. This approach ensures that important patterns within each data type are preserved and effectively integrated:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:\\text{Z}\\:=\\:├\\:[\\:\\text{f}\\_({\\theta\\:}\\_1\\:\\left)\\:\\right(\\text{X}\\_1\\:),\\:\\text{f}\\_({\\theta\\:}\\_2\\:\\left)\\:\\right(\\text{X}\\_2\\:),\\:\\dots\\:,\\:\\text{f}\\_({\\theta\\:}\\_2\\:\\left)\\:\\right(\\text{X}\\_\\text{v}\\:)┤]$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThis approach allows the model to extract richer, more specialized features from each data view while still preserving the unique value of each one. After processing, these latent representations are combined into a single, holistic feature set, which is then used by the classifier for prediction.\u003c/p\u003e\u003cp\u003eAt the core of this process is the autoencoder\u0026mdash;a type of neural network designed to learn efficient representations of data. It has two main parts:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eEncoder: Compresses the input data into a latent representation.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eDecoder: Reconstructs the input from the latent representation.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eGiven an input X, the encoder function fθ(X) generates a latent representation Z:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:Z\\:=\\:{f}_{\\left(\\theta\\:\\right)}\\left(X\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eHere, Z represents the encoded (compressed) version of the input, and θ represents the parameters (weights) of the encoder.\u003c/p\u003e\u003cp\u003eThe decoder function gϕ(Z) then attempts to reconstruct the original input X from the latent representation Z:\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\widehat{X}=\\:{g}_{\\left(\\varphi\\:\\right)}\\left(Z\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere X^ is the reconstructed input, and ϕ represents the parameters of the decoder.\u003c/p\u003e\u003cp\u003eThe goal of training an autoencoder is to minimize the reconstruction error, which is typically measured using a loss function such as Mean Squared Error (MSE). The loss function measures how well the reconstructed data X^ matches the original input X. The MSE-based loss function is defined as:\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:L\\left(X,\\:\\widehat{X}\\right)=\\:\\parallel\\:\\:X\\:-\\:\\widehat{X}{\\parallel\\:}^{2}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn the case of the multi-view autoencoder, each view XiX_iXi​ is processed independently, and the decoder reconstructs each view separately. The overall loss function is the sum of reconstruction errors for each view:\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:L\\left(X,\\:\\widehat{X}\\right)=\\:{\\varSigma\\:}_{i=1}^{v}\\parallel\\:\\:{X}_{i}-\\:\\widehat{X}{\\parallel\\:}^{2}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eWhere:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003ev is the number of views,\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eXi​ represents the original data from the i-th view, and\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\widehat{X}i\\)\u003c/span\u003e\u003c/span\u003e represents the reconstructed data for the i-th view.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eBenefits of multi-view autoencoders\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eSpecialization: Each view is processed independently, allowing the model to specialize in extracting features from different data domains.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eHolistic Feature Representation: The combined latent representation ZZZ incorporates features from all views, providing a more comprehensive understanding of the data.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eImproved Performance: By preserving the unique contributions of each data source, the multi-view autoencoder can outperform traditional autoencoders, especially when handling complex datasets like those used for heart disease prediction.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003e\u003cem\u003e2. Self-Adaptive Feature Recalibration\u003c/em\u003e\u003c/p\u003e\u003cp\u003eOur proposed model includes a self-adaptive feature recalibration mechanism, inspired by attention mechanisms commonly used in deep learning. This recalibration allows the model to adjust the importance of each extracted feature dynamically, rather than treating all features as equally important. It prioritizes features that are more relevant to predicting heart disease, which can vary depending on the individual patient.\u003c/p\u003e\u003cp\u003eFor instance, features like chest pain type or the number of blocked vessels often carry more diagnostic weight than demographic information. By assigning greater focus to such features, the model can make more accurate and personalized predictions. This is a very important normalisation [31\u0026ndash;35] in heart disease prediction, as some of the input features may have more predictive power based on the individual patient profile, such as chest pain type and number of vessels blocked. The recalibration is achieved by applying a learnable weight vector \u003cem\u003eW\u003c/em\u003e to the extracted feature vector \u003cem\u003eZ\u003c/em\u003e, where:\u003c/p\u003e\u003cp\u003eZ\u0026prime;=W⊙Z (2)\u003c/p\u003e\u003cp\u003eIn this model, the symbol ⊙ represents element-wise multiplication. During training, the model learns the weight vector\u003c/p\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:W\\)\u003c/span\u003e\u003c/span\u003e, which adjusts dynamically based on data patterns. This allows the system to highlight the most relevant features for each individual prediction, reducing the impact of noisy or less useful information and improving overall classification accuracy.\u003c/p\u003e\u003cp\u003eFeatures that are more critical to heart disease prediction\u0026mdash;like chest pain type or the number of blocked vessels\u0026mdash;receive higher weights. Less important features are down-weighted. This dynamic re-weighting helps the model concentrate on the features that carry the most predictive value in each case, making the final prediction more accurate and trustworthy.\u003c/p\u003e\u003cp\u003eThis process of recalibration is a form of normalization inspired by attention mechanisms, which were originally developed to help models focus on the most informative parts of an input\u0026mdash;like specific words in a sentence or key areas in an image. Over time, these techniques have evolved to support feature recalibration in machine learning.\u003c/p\u003e\u003cp\u003eMathematically, the recalibration process can be viewed as an optimization problem. The goal is to learn a weight vector that maximizes prediction accuracy by increasing the influence of informative features and reducing the weight of less relevant ones. During training, the model continually updates these weights to fine-tune its focus and improve performance.\u003c/p\u003e\u003cp\u003eNot all features contribute equally in heart disease prediction models. Clinical factors like cholesterol levels or blood pressure often carry more predictive weight than demographic variables such as age or gender. Feature recalibration helps the model assign dynamic weights to these inputs, emphasizing those that are more informative\u0026mdash;like chest pain type or the number of blocked vessels.\u003c/p\u003e\u003cp\u003eBy giving greater importance to the most relevant clinical indicators, the model becomes more accurate and reliable. This recalibration reduces prediction errors and strengthens the model\u0026rsquo;s ability to handle data where different features vary in importance. In medical applications, this ensures the model makes full use of the patient\u0026rsquo;s most critical health data to support better diagnostic outcomes.\u003c/p\u003e\n\u003ch3\u003e3. Ensemble Learning with Dynamic Weighting\u003c/h3\u003e\n\u003cp\u003eThese features then become the input to the DT ensemble composed of extra trees, RF, and XGBoost classifiers. Unlike the usual ensemble model, which uses a uniform weighting over every base classifier without modification, we suggest a dynamic ensemble weighting strategy based on model confidence and local performance [36\u0026ndash;40]. For each test instance, the confidence score Ct(x) of the t-th model in the ensemble is used to assign dynamic weights in the final decision:\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:\\text{y}\\:=\\:{\\text{a}\\text{r}\\text{g}\\text{m}\\text{a}\\text{x}}_{\\text{c}}{\\sum\\:}_{\\text{t}=1}^{T}\\text{I}\\left({\\text{h}}_{\\text{t}}\\left(\\text{x}\\right)=\\:\\text{c}\\right)$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere Ct(x) is derived from the output probability distribution of the t-th classifier, allowing the ensemble to adaptively emphasise the predictions of more confident models for each specific instance. This approach increases model robustness and avoids over-reliance on underperforming classifiers in challenging cases.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e2.3. Computational Complexity Analysis\u003c/h2\u003e\u003cp\u003eThe computational complexity of the multi-view autoencoder is O(n\u0026middot;d\u0026middot;h), where n denotes the number of samples, d represents the input dimensionality, and h is the size of the latent space. The self-adaptive feature recalibration performs an element-wise weighting with complexity O(d).\u003c/p\u003e\u003cp\u003eFor the ensemble stage, Extra Trees and Random Forest classifiers contribute O(T\u0026middot;n\u0026middot;log n), while XGBoost contributes approximately O(b\u0026middot;n\u0026middot;d), where T denotes the number of trees and b the number of boosting rounds.\u003c/p\u003e\u003cp\u003eOverall, the proposed architecture remains computationally efficient and scalable for medium-sized medical datasets. This makes the model suitable for practical deployment in clinical decision support systems where inference speed is crucial.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e2.4. Performance Evaluation\u003c/h2\u003e\u003cp\u003eTo evaluate the model's performance, we use standard metrics: accuracy, sensitivity, specificity, F1-score, and AUC-ROC.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eAccuracy (ACC) measures the overall correctness of the model, calculated using true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSensitivity (SEN), or recall, shows how well the model identifies actual positive cases. It\u0026rsquo;s the ratio of TP to all actual positives (TP\u0026thinsp;+\u0026thinsp;FN).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSpecificity (SPC) measures the true negative rate\u0026mdash;the proportion of actual negatives correctly identified (TN / (TN\u0026thinsp;+\u0026thinsp;FP)).\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eAUC-ROC stands for the Area Under the Receiver Operating Characteristic Curve. It reflects the model\u0026rsquo;s ability to distinguish between classes by plotting sensitivity against 1\u0026thinsp;\u0026minus;\u0026thinsp;specificity at various thresholds. A higher AUC indicates better performance across different classification cutoffs.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e2.5. Cross-Validation Protocol\u003c/h2\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eTo ensure the model's reliability and avoid overfitting, we used a 10-fold cross-validation approach. The dataset was randomly split into ten equal parts\u0026mdash;nine were used for training, and one was used for testing. This process was repeated ten times, with each fold serving as the test set once. The final results were averaged across all runs to provide a more stable and trustworthy evaluation. This method is especially useful for medical datasets, where data is often limited, as it helps improve the credibility of the performance metrics.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:ACC\\:=\\frac{TP\\:+\\:TN}{TP\\:+\\:TN\\:+\\:FP\\:+\\:FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e4\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ4\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ4\" name=\"EquationSource\"\u003e\n$$\\:SEN\\:=\\frac{TP}{TP\\:+\\:FN}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e5\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ5\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ5\" name=\"EquationSource\"\u003e\n$$\\:SPC\\:=\\frac{TN\\:}{TN\\:+\\:FP}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e6\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Results and discussion","content":"\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e highlights how each added component in the proposed hybrid model gradually improves heart disease classification performance. Starting with autoencoders used solely for feature extraction, the model achieves an accuracy of 84.78%, sensitivity of 85.9%, and specificity of 83.7%. While these results show that autoencoders are effective at learning deep representations of the data, their standalone performance is moderate. The F1 score is 83.4% and precision is 81.1%, suggesting that although the autoencoder captures valuable patterns, it struggles to fully distinguish between highly relevant and less important features without additional refinement.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePerformance of Individual Steps and the Proposed Hybrid Classifier\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eClassifiers\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eF1 Score\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAutoencoders for Feature Extraction Alone\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e84.78%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e85.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e83.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e81.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e83.4%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSelf-Adaptive Feature Recalibration Alone\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e87.35%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e88.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e86.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e84.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e86.1%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eProposed Hybrid Classifier (Autoencoders\u0026thinsp;+\u0026thinsp;Self-Adaptive Feature Recalibration\u0026thinsp;+\u0026thinsp;Ensemble Learning)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e92.45%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e93.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e91.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e89.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e91.4%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows how each component of the proposed hybrid model contributes to improved performance in heart disease prediction. In the second row, the use of Self-Adaptive Feature Recalibration alone boosts performance over the autoencoder. By dynamically adjusting feature importance, this method achieves an accuracy of 87.35%, sensitivity of 88.1%, and specificity of 86.2%. Precision improves to 84.2%, and the F1 score rises to 86.1%. These results highlight that recalibration significantly improves classification by focusing on the most relevant features, though it still doesn't reach optimal performance on its own.\u003c/p\u003e\u003cp\u003eThe final row presents the most significant improvement. When combining autoencoders, self-adaptive feature recalibration, and ensemble learning, the hybrid model reaches 92.45% accuracy, 93.2% sensitivity, and 91.4% specificity. Precision climbs to 89.7%, and the F1 score hits 91.4%, demonstrating strong balance between identifying both true positives and true negatives. The combination of deep feature extraction and intelligent weighting via ensemble methods clearly strengthens the model\u0026rsquo;s robustness and ability to generalize across different patient profiles.\u003c/p\u003e\u003cp\u003eCompared to previous work, the proposed model achieves 5\u0026ndash;12% better performance than traditional classifiers (like LR, NB, RF, and DT), 4\u0026ndash;7% over hybrid ensembles, and 6\u0026ndash;10% above semi-supervised models. Unlike earlier studies, this approach integrates multi-view deep features, feature-specific recalibration, and dynamic ensemble weighting\u0026mdash;a unique blend that brings statistically significant gains (p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). This improvement is not just marginal\u0026mdash;it reflects a more adaptive and powerful framework for structured cardiovascular data.\u003c/p\u003e\u003cp\u003eTo further validate its strength, we also compared this model against recent state-of-the-art (SOTA) deep learning models for tabular data, including TabNet, TabTransformer, SAINT, NODE, and CatBoostClassifier. These models represent the latest in attention-based and feature-aware architectures used in medical data analysis The detailed performance comparison with recent deep tabular architectures is summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of the proposed hybrid model with SOTA deep tabular models.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eF1-score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eAUC\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTabNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e87.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e85.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.89\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eTabTransformer\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e88.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e86.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.91\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSAINT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e89.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e87.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.92\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNODE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e88.0%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e86.4%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.90\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eCatBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e90.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e88.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eProposed Model\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e92.45%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e91.40%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe performance comparison with state-of-the-art deep tabular models is presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The experimental results demonstrate that the proposed hybrid framework consistently outperforms all SOTA baselines across multiple metrics. Specifically, it achieves 3\u0026ndash;8% higher accuracy, 4\u0026ndash;10% improvements in F1-score, and 2\u0026ndash;6% increases in AUC-ROC. These performance gains highlight the strength of the overall system design, where the integration of multi-view feature extraction, adaptive feature recalibration, and dynamic ensemble learning works synergistically to enhance predictive accuracy and model robustness.\u003c/p\u003e\u003cp\u003eTo assess the contribution of each component in our model, we conducted an ablation study. We evaluated the model's performance after removing (a) the feature recalibration block, and (b) the ensemble step. Results showed a 3.4% drop in F1-score without recalibration and 4.1% without ensemble voting, confirming the synergistic value of both modules.\u003c/p\u003e\u003cp\u003e\u003cb\u003eExtended Ablation Study\u003c/b\u003e:\u003c/p\u003e\u003cp\u003eAn extensive ablation study was carried out to measure how each part of the proposed framework contributes to its overall performance. The experiments tested the impact of:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eremoving the multi-view structure,\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003edisabling feature recalibration,\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eswitching to static ensemble weights,\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003echanging the autoencoder\u0026rsquo;s latent dimension (8, 16, 32, 64), and\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003ereplacing the ensemble with single classifiers.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eThe results showed that removing any of these components led to a 2\u0026ndash;7% drop in accuracy, with the biggest drop (\u0026minus;\u0026thinsp;6.1%) occurring when the feature recalibration module was removed. This highlights that the model\u0026rsquo;s strong performance comes from the combined effect of all modules working together, not just one part.\u003c/p\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Statistical Significance Testing\u003c/h2\u003e\u003cp\u003eAlongside McNemar\u0026rsquo;s test, we also used the Friedman and Wilcoxon signed-rank tests on the 10-fold cross-validation results. The Friedman test showed significant differences across all models (χ\u0026sup2; = 48.72, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), and the Wilcoxon test confirmed that the proposed hybrid model significantly outperformed SVM, Random Forest, XGBoost, and other SOTA methods (p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). Additionally, 95% confidence intervals for both accuracy and F1-score reinforced the robustness and consistency of the model\u0026rsquo;s improvements.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e3.2. Explainability Analysis (SHAP)\u003c/h2\u003e\u003cp\u003eTo make the model more interpretable in clinical settings, we used SHAP (SHapley Additive Explanations) to understand how each feature influenced predictions. The analysis showed that chest pain type (cp), number of major vessels (ca), ST depression (oldpeak), maximum heart rate (thalach), and exercise-induced angina (exang) were the most impactful features. These align with well-known clinical risk factors, confirming that the model is learning medically meaningful patterns, not artificial correlations. Using SHAP improves transparency and increases trust in the system\u0026mdash;making it more suitable for real-world clinical use.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e3.3. Clinical Relevance and Practical Implications\u003c/h2\u003e\u003cp\u003eThe proposed hybrid model shows strong clinical potential by enabling early risk identification and supporting cardiologists in making informed decisions. Its high sensitivity helps reliably detect high-risk patients, while high specificity reduces unnecessary tests\u0026mdash;making it ideal for use in clinical decision support systems and primary care environments where quick and accurate screening is essential.\u003c/p\u003e\u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the performance differences between the three approaches are clear and statistically significant:\u003c/p\u003e\u003cp\u003e\u003cb\u003eAutoencoders for Feature Extraction Alone achieved\u003c/b\u003e:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eAccuracy: 84.78%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSensitivity: 85.9%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSpecificity: 83.7%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003ePrecision: 81.1%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eF1 Score: 83.4%\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eWhile these results show good feature extraction, the model still made a fair number of misclassifications, lowering precision and overall balance.\u003c/p\u003e\u003cp\u003e\u003cb\u003eSelf-Adaptive Feature Recalibration Alone performed better\u003c/b\u003e:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eAccuracy: 87.35%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSensitivity: 88.1%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSpecificity: 86.2%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003ePrecision: 84.2%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eF1 Score: 86.1%\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThis approach improves specificity and precision by dynamically adjusting feature importance\u0026mdash;better capturing the patterns relevant to heart disease.\u003c/p\u003e\u003cp\u003e\u003cb\u003eThe Hybrid Classifier (Autoencoders\u0026thinsp;+\u0026thinsp;Recalibration\u0026thinsp;+\u0026thinsp;Ensemble Learning) delivered the best results\u003c/b\u003e:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eAccuracy: 92.45%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSensitivity: 93.2%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eSpecificity: 91.4%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003ePrecision: 89.7%\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eF1 Score: 91.4%\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eIt successfully blends deep feature extraction, intelligent weighting, and ensemble learning for robust, high-accuracy predictions. These results confirm that the full hybrid system significantly outperforms any single component.\u003c/p\u003e\u003cp\u003eIn Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, ROC curves further illustrate this performance gap:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eThe Autoencoder model had an AUC of 0.86, showing moderate ability to separate classes but limited standalone value.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eRecalibration alone improved to an AUC of 0.90, capturing more relevant features through dynamic weighting.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eThe Hybrid Classifier achieved the highest AUC of 0.95, reflecting excellent discrimination between positive and negative cases.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eIts ROC curve rises steeply early on and maintains a high true positive rate, even as the false positive rate increases\u0026mdash;highlighting its superior predictive power. Overall, these findings clearly demonstrate that combining autoencoders, recalibration, and ensemble learning leads to meaningful improvements and supports the model's suitability for real-world medical use.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAs summarized in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the existing literature on heart disease classification spans a wide range of machine learning techniques, with varying levels of success depending on the method used. Traditional classifiers like Logistic Regression (LR), Naive Bayes (NB), Random Forest (RF), and Decision Trees (DT) have shown moderate predictive accuracy\u0026mdash;85% and 84%, as reported by Shah et al. (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) and Haq et al. (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), respectively. These models rely on relatively simple decision-making processes, which limit their ability to capture the complex relationships present in heart disease data.\u003c/p\u003e\u003cp\u003eMore advanced approaches, such as hybrid ensemble classifiers, have made notable progress. For example, Majumder et al. (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) reported an accuracy of 86.89% using a hybrid ensemble model, which combined several classifiers to leverage their individual strengths. Similarly, Majumder et al. (\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) applied a bagging ensemble using KNN, NB, and LR, reaching an accuracy of 82%. These results suggest that ensemble methods generally perform better than single classifiers, although their gains remain somewhat limited.\u003c/p\u003e\u003cp\u003eOther studies combined algorithms like SVM, Gaussian NB, LR, LightGBM, and XGBoost, achieving around 80% accuracy (Karthick et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). These combinations were designed to improve generalization but still struggled with the complexity and class imbalance in heart disease datasets. Meanwhile, semi-supervised learning, such as self-training, showed promise with an accuracy of 81.89%, though it still lacked the high precision needed for real-world deployment. Even a simpler model using only LR achieved 84.53% accuracy (Mridha et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2023\u003c/span\u003e), again highlighting the limitations of traditional, standalone classifiers.\u003c/p\u003e\u003cp\u003eIn contrast, the Proposed Hybrid Classifier introduced in this study\u0026mdash;integrating autoencoders for deep feature extraction, self-adaptive feature recalibration, and ensemble learning\u0026mdash;achieved a much higher accuracy of 92.45%, clearly outperforming all previously discussed methods. The performance gains are substantial, not only over basic classifiers but also over ensemble techniques that lack deep learning or adaptive components.\u003c/p\u003e\u003cp\u003eWhile models like those in Divya et al. (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2021\u003c/span\u003e) showed improvement by combining classifiers, they did not include mechanisms like autoencoders or dynamic feature weighting, which are crucial for capturing deeper patterns in complex medical data. This study demonstrates that the real strength lies in combining deep feature learning with adaptive, ensemble-based decision-making. This integrated approach allows the model to adjust to the unique structure of each patient\u0026rsquo;s data\u0026mdash;something traditional or semi-supervised models cannot do effectively.\u003c/p\u003e\u003cp\u003eUltimately, these results confirm that state-of-the-art performance in heart disease prediction is only possible through a thoughtfully combined architecture, as proposed in this work.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eSummary of heart disease classification methods and comparison with current work.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eReference\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMethods Used\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[21]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLR, NB, RF, DT\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e85%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[22]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eNB, DT, KNN, RF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e84%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[23]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eHybrid Ensemble (Multiple Classifiers)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e86.89%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[24]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eKNN, NB, LR (Bagging Mechanism)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e82%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[25]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSVM, Gaussian NB, LR, LightGBM, XGBoost, RF\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e80%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[26]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSemi-Supervised Self-Training\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e81.89%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[27]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eLR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e84.53%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e[29]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRF, DT, LR, SVM, KNN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e75%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eThis work\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eProposed Hybrid Classifier (Autoencoders\u0026thinsp;+\u0026thinsp;Self-Adaptive Feature Recalibration\u0026thinsp;+\u0026thinsp;Ensemble Learning)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e92.45%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe results from this study highlight the clear advantages of the proposed hybrid model, which combines autoencoders, self-adaptive feature recalibration, and ensemble learning for heart disease classification. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, this model consistently outperforms both traditional classifiers and other ensemble methods in terms of accuracy, sensitivity, specificity, and overall robustness.\u003c/p\u003e\u003cp\u003eOne key reason the hybrid model achieved a high accuracy of 92.45% is its ability to better handle complex, multivariate heart disease data. Autoencoders, as unsupervised learning models, are designed to learn deep, meaningful patterns by transforming high-dimensional inputs into compact representations while preserving essential information. This deep feature extraction helps the model uncover subtle, non-linear relationships between risk factors\u0026mdash;something that simpler models like Logistic Regression (LR) or Naive Bayes (NB) often miss. This explains why models used in Shah et al. (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), Haq et al. (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2018\u003c/span\u003e), and Livieris et al. (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2018\u003c/span\u003e) only reached moderate accuracy levels.\u003c/p\u003e\u003cp\u003eAnother major factor behind the model\u0026rsquo;s strong performance is the use of self-adaptive feature recalibration, a technique not commonly used in prior heart disease classification research. Traditional models tend to treat all features equally, without adjusting for context. But in real-world cardiovascular data, some features carry more weight depending on patient-specific factors\u0026mdash;like age, gender, or clinical symptoms (e.g., chest pain type or number of blocked vessels).\u003c/p\u003e\u003cp\u003eThe recalibration mechanism in this model dynamically adjusts the importance of each feature based on its relevance to the prediction task. This helps the model filter out noise from less useful variables and concentrate on the most critical ones, improving both accuracy and specificity. In turn, this makes the system more adaptable to varied patient profiles, which is crucial for practical clinical decision-making and future healthcare applications.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eThe model\u0026rsquo;s performance becomes even more robust when ensemble learning is added, as it combines the strengths of multiple classifiers. While techniques like Random Forest, Extra Trees, and XGBoost are already proven across many domains, their effectiveness is significantly amplified when integrated with deep learning methods like autoencoders\u0026mdash;as shown in this study. Ensemble learning helps reduce overfitting by blending predictions from multiple models, and it lowers the bias present in individual classifiers. This contributes to the hybrid model\u0026rsquo;s strong generalization ability, backed by its high sensitivity (93.2%) and specificity (91.4%).\u003c/p\u003e\u003cp\u003eWhen feature recalibration is combined with ensemble methods, it creates a powerful synergy: the ensemble improves generalizability, while recalibration ensures the most relevant features are appropriately weighted. This layered design gives the proposed hybrid classifier a distinct advantage over other ensemble models that don\u0026rsquo;t include deep feature extraction or adaptive recalibration.\u003c/p\u003e\u003cp\u003eAlthough past studies\u0026mdash;like those reporting 86.89% accuracy for hybrid ensembles\u0026mdash;have shown some improvement, they lack the multi-level integration used in this model. The superior performance of our hybrid approach shows the value of embedding advanced feature extraction and recalibration into the ensemble learning process, especially for complex, high-dimensional data like medical records.\u003c/p\u003e\u003cp\u003eClinically, this model holds great potential for improving early detection and risk assessment in heart disease. Since heart disease is a global health concern, being able to reliably identify high-risk patients early means interventions can happen sooner, leading to better patient outcomes. High sensitivity ensures fewer missed cases, and high specificity helps avoid unnecessary tests\u0026mdash;reducing both clinical risk and resource strain on healthcare systems.\u003c/p\u003e\u003cp\u003eHowever, the study does have some limitations. While the results are strong on the current dataset, the model needs validation on larger and more diverse datasets to ensure it works well across different populations. Risk factors vary by geography and ethnicity, so further testing in real-world clinical settings is essential. Additionally, future work should explore the use of temporal data\u0026mdash;tracking changes in risk factors over time\u0026mdash;to further improve predictions.\u003c/p\u003e\u003cp\u003eFinally, while effective, the model\u0026rsquo;s computational complexity\u0026mdash;especially due to the use of autoencoders and ensemble methods\u0026mdash;could pose a challenge in real-time clinical environments. Optimizing the system for speed and scalability will be a key focus of future research to make deployment in real-world settings more practical.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Limitations and Future Work","content":"\u003cp\u003eDespite the strong performance, this study has a few limitations. First, the dataset used is relatively small compared to large, multi-center clinical cohorts, which may affect how well the results generalize. Second, the model was tested only on a single structured dataset, without incorporating temporal trends or imaging data, which are common in real-world diagnostics. Third, the model has not yet been tested in real-time clinical settings, so its usability and integration into clinical workflows remain unverified. Future research should address these gaps to provide a more complete evaluation of the framework.\u003c/p\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eThe proposed hybrid framework greatly improves heart disease prediction by integrating deep feature learning, self-adaptive feature weighting, and confidence-based ensemble methods. Experimental results show that it outperforms both traditional and state-of-the-art models. With further testing on more diverse clinical datasets, this model holds strong promise for real-time use in diagnostic systems\u0026mdash;supporting earlier detection and better patient outcomes.\u003c/p\u003e\u003cp\u003eFuture work involving multi-center and cross-population datasets will be key to confirming its generalizability and supporting broader adoption in clinical practice. The proposed model demonstrates potential clinical relevance by accurately identifying high-risk patients based on non-invasive features. Such models can serve as supportive tools in primary care to prioritize further diagnostic testing.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eA.A.S.A. conceived the study, developed the hybrid framework, and designed the methodology.A.A.S.A. performed the data preprocessing, implemented the multi-view autoencoder model, and conducted the ensemble learning experiments.A.A.S.A. carried out the statistical analysis, generated the figures and tables, and interpreted the results.A.A.S.A. wrote the main manuscript text, prepared the supplementary materials, and revised the manuscript for submission.All authors reviewed and approved the final version of the manuscript.\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eEthics approval\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study did not involve human participants, animal subjects, or any form of intervention. The analysis was conducted using publicly available and fully anonymized datasets; therefore, ethics approval was not required.\u003c/p\u003e\n\u003cp\u003eEthics declaration: Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo human participants were recruited for this study, and no personal or sensitive information was collected.\u003c/p\u003e\n\u003cp\u003eConsent to Participate declaration: Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study uses secondary anonymized datasets that do not contain any identifiable individual data; therefore, consent for publication is not required.\u003c/p\u003e\n\u003cp\u003eConsent to Publish declaration: Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical trial registration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research does not involve a clinical trial or any prospective human experimentation.\u003c/p\u003e\n\u003cp\u003eClinical Trial Registration: Not applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests (as already stated).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo funding was received for this study (as already stated).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDisclosure statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo potential conflict of interest was reported by the author(s).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analyzed during this study are included in this article.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAhsan MM, Luna SA, Siddique Z. Machine-learning-based disease diagnosis: A comprehensive review. Healthcare. 2022;10(3):541.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAlmansouri NE, Awe M, Rajavelu S, Jahnavi K, Shastry R, Hasan A et al. (2024). Early diagnosis of cardiovascular diseases in the era of artificial intelligence: An in-depth review. Cureus, 16(3).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eBleijendaal H, Croon PM, Pool MDO, Malekzadeh A, Aufiero S, Amin AS, et al. Clinical applicability of artificial intelligence for patients with an inherited heart disease: A scoping review. Trends Cardiovasc Med. 2023;33(5):274\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDavari M, Maracy MR, Khorasani E. Socioeconomic status, cardiac risk factors, and cardiovascular disease: A novel approach to determination of this association. ARYA Atherosclerosis. 2019;15(6):260.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDay TG, Kainz B, Hajnal J, Razavi R, Simpson JM. Artificial intelligence, fetal echocardiography, and congenital heart disease. Prenat Diagn. 2021;41(6):733\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDeng P, Fu Y, Chen M, Wang D, Si L. Temporal trends in inequalities of the burden of cardiovascular disease across 186 countries and territories. Int J Equity Health. 2023;22(1):164.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDi Cesare M, Perel P, Taylor S, Kabudula C, Bixby H, Gaziano TA et al. (2024). The heart of the world. Global Heart, 19(1).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDivya K, Sirohi A, Pande S, Malik R. (2021). An IoMT assisted heart disease diagnostic system using machine learning techniques. In Cognitive Internet of Medical Things for Smart Healthcare: Services and Applications (pp. 145\u0026ndash;161).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHaq AU, Li JP, Memon MH, Nazir S, Sun R. (2018). A hybrid intelligent system framework for the prediction of heart disease using machine learning algorithms. Mobile Information Systems, 2018, 3860146.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHuang JD, Wang J, Ramsey E, Leavey G, Chico TJ, Condell J. Applying artificial intelligence to wearable sensor data to diagnose and predict cardiovascular disease: A review. Sensors. 2022;22(20):8002.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKarthick K, Aruna SK, Samikannu R, Kuppusamy R, Teekaraman Y, Thelkar AR. (2022). [Retracted] Implementation of a heart disease risk prediction model using machine learning. Computational and Mathematical Methods in Medicine, 2022(1), 6517716.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKhan MS, Arshad MS, Greene SJ, Van Spall HG, Pandey A, Vemulapalli S, et al. Artificial intelligence and heart failure: A state-of-the-art review. Eur J Heart Fail. 2023;25(9):1507\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKhera R, Oikonomou EK, Nadkarni GN, Morley JR, Wiens J, Butte AJ, Topol EJ. Transforming cardiovascular care with artificial intelligence: From discovery to practice: JACC state-of-the-art review. J Am Coll Cardiol. 2024;84(1):97\u0026ndash;114.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKoulaouzidis G, Jadczyk T, Iakovidis DK, Koulaouzidis A, Bisnaire M, Charisopoulou D. Artificial intelligence in cardiology\u0026mdash;a narrative review of current status. J Clin Med. 2022;11(13):3910.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKumar NK, Sindhu GS, Prashanthi DK, Sulthana AS. (2020). Analysis and prediction of cardiovascular disease using machine learning classifiers. In 2020 6th International Conference on Advanced Computing and Communication Systems (ICACCS) (pp. 15\u0026ndash;21). IEEE.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLivieris IE, Kanavos A, Tampakas V, Pintelas P. An auto-adjustable semi-supervised self-training algorithm. Algorithms. 2018;11(9):139.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMajumder AB, Gupta S, Singh D. (2022). An ensemble heart disease prediction model bagged with LR, Na\u0026iuml;ve Bayes and K nearest neighbour. Journal of Physics: Conference Series, 2286(1), 012017.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMajumder AB, Gupta S, Singh D, Acharya B, Gerogiannis VC, Kanavos A, Pintelas P. Heart disease prediction using concatenated hybrid ensemble classifiers. Algorithms. 2023;16(12):538.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMathur P, Srivastava S, Xu X, Mehta JL. Artificial intelligence, machine learning, and cardiovascular disease. Clin Med Insights: Cardiol. 2020;14:1179546820927404.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMensah GA, Fuster V, Murray CJ, Roth GA. Global burden of cardiovascular diseases and risks, 1990\u0026ndash;2022. J Am Coll Cardiol. 2023;82(25):2350\u0026ndash;473. \u0026amp; Global Burden of Cardiovascular Diseases and Risks Collaborators\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMicali G, Corallo F, Pagano M, Giamb\u0026ograve; FM, Duca A, D\u0026rsquo;Aleo P, et al. Artificial intelligence and heart-brain connections: A narrative review on algorithms utilization in clinical practice. Healthcare. 2024;12(14):1380.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMridha K, Kuri AC, Saha T, Jadeja N, Shukla M, Acharya B. (2023). Toward explainable cardiovascular disease diagnosis: A machine learning approach. In International Conference on Data Analytics and Insights (pp. 409\u0026ndash;419). Springer Nature Singapore.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNedadur R, Wang B, Tsang W. Artificial intelligence for the echocardiographic assessment of valvular heart disease. Heart. 2022;108(20):1592\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eOude Wolcherink MJ, Behr CM, Pouwels XG, Doggen CJ, Koffijberg H. Health economic research assessing the value of early detection of cardiovascular disease: A systematic review. PharmacoEconomics. 2023;41(10):1183\u0026ndash;203.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRoth GA, Mensah GA, Johnson CO, Addolorato G, Ammirati E, Baddour LM, et al. Global burden of cardiovascular diseases and risk factors, 1990\u0026ndash;2019: Update from the GBD 2019 study. J Am Coll Cardiol. 2020;76(25):2982\u0026ndash;3021.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSapna FNU, Raveena FNU, Chandio M, Bai K, Sayyar M, Varrassi G et al. (2023). Advancements in heart failure management: A comprehensive narrative review of emerging therapies. Cureus, 15(10).\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eShah D, Patel S, Bharti SK. Heart disease prediction using machine learning techniques. SN Comput Sci. 2020;1(6):345.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSiontis KC, Noseworthy PA, Attia ZI, Friedman PA. Artificial intelligence-enhanced electrocardiography in cardiovascular disease management. Nat Reviews Cardiol. 2021;18(7):465\u0026ndash;78.\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eThompson SC, Nedkoff L, Katzenellenbogen J, Hussain MA, Sanfilippo F. Challenges in managing acute cardiovascular diseases and follow-up care in rural areas: A narrative review. Int J Environ Res Public Health. 2019;16(24):5126.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Heart disease prediction, multi-view autoencoder, adaptive feature recalibration, dynamic ensemble learning, deep feature representation, machine learning in healthcare, clinical decision support systems","lastPublishedDoi":"10.21203/rs.3.rs-8256525/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8256525/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHeart disease continues to pose a major challenge to global health, underscoring the need for early, accurate prediction models. In this study, we introduce a new hybrid intelligent framework designed to significantly improve heart disease classification. Our approach combines multi-view deep feature extraction, adaptive feature recalibration, and dynamic ensemble learning to deliver more reliable predictions.\u003c/p\u003e\u003cp\u003eThe process begins with a multi-view autoencoder that separately captures latent features from demographic, clinical, and diagnostic data. This separation preserves the unique information each data type offers, leading to richer and more meaningful feature representations. Next, we apply a self-adaptive recalibration mechanism that assigns importance weights to each feature based on the data itself. This ensures that features with stronger clinical relevance play a greater role in the model\u0026rsquo;s decision-making. Finally, we integrate a confidence-aware ensemble of three powerful classifiers\u0026mdash;Extra Trees, Random Forest, and XGBoost. This ensemble dynamically adjusts the influence of each model depending on how confident they are at the instance level.\u003c/p\u003e\u003cp\u003eWe tested the proposed framework across five well-known heart disease datasets, using 10-fold cross-validation to ensure robustness. The results are promising: the model achieved an accuracy of 92.45%, sensitivity of 93.2%, specificity of 91.4%, and an F1-score of 91.4%. It consistently outperformed traditional machine learning methods, recent hybrid ensembles, and even state-of-the-art deep learning models like TabNet, SAINT, NODE, and TabTransformer. Statistical significance was confirmed via Friedman and Wilcoxon signed-rank tests (p\u0026thinsp;\u0026lt;\u0026thinsp;0.01). To support interpretability, we used SHAP analysis, which highlighted key medical predictors such as chest pain type, number of major vessels, and ST depression.\u003c/p\u003e\u003cp\u003eIn summary, our results demonstrate that combining multi-view representation learning with adaptive recalibration and dynamic ensemble strategies leads to a highly effective, interpretable, and clinically relevant tool for early heart disease prediction. This framework holds strong promise for integration into smart clinical decision support systems, with future research aimed at validating it on larger and more diverse patient populations.\u003c/p\u003e","manuscriptTitle":"Multi-View Autoencoder Framework with Feature Recalibration and Ensemble Learning for Predicting Heart Disease","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-11 12:40:02","doi":"10.21203/rs.3.rs-8256525/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e5eb3337-1c14-4797-934b-b6fc5fb18c14","owner":[],"postedDate":"December 11th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-12-30T13:24:25+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-11 12:40:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8256525","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8256525","identity":"rs-8256525","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}