A Modality-Specific Deep Learning Framework for Cardiac Disorder Screening Using Electrocardiogram Signals and Clinical Data

With an estimated 17.9 million deaths occurring yearly, representing 32% of all deaths worldwide [1], cardiovascular diseases are currently the leading killers worldwide. Among them are cardiac arrhythmias and coronary artery disease (CAD), the most common and high- risk clinical forms of CVDs, each of which can give rise to life-threatening conditions if not found and treated in time. Arrhythmias are abnormal heart rhythms generated by irregular electrical activity in the heart muscle, while in CAD, the coronary arteries become narrowed or blocked, usually because of atherosclerosis [2]. Such conditions have to be diagnosed early to enable timely intervention and further management, but the conventional ways of performing this, such as manual electrocardiogram (ECG) interpretation and invasive coronary angiography, are time-consuming, require clinical expertise, and are at times subject to human error through observer variability.

Recent developments in Artificial Intelligence (AI) and machine learning have transformed the arena of healthcare analytics, especially in automated medical diagnostics. Since patient records and biosignals have been increasingly digitized, deep learning models have become effective tools to extract complex patterns from high-dimensional data, performing much better than traditional machine learning methods under many clinical circumstances [3, 4]. Specifically, Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), particularly those based on Long Short-Term Memory (LSTM), have shown particular promise in the analysis of ECG signals, a primary time-series data type used in cardiac care [5]. In contrast, Deep Neural Networks (DNNs) are more apt to deal with structured clinical data, where complex relationships among risk factors, laboratory variables, and disease outcomes can be learned.

Deep learning methods for cardiac disease detection have been the subject of a few studies. For instance, a 34-layer CNN model named Cardiologist-Level Arrhythmia Detection developed by Hannun et al. [6] displayed expert-level performance on PhysioNet data. In the same manner, Yildirim et al. [7] showed that a deep bidirectional LSTM-CNN can be used for multi-class arrhythmia classification, attaining excellent accuracy on benchmark ECG datasets. Concerning CAD prediction, a study by Khosla et al. [8] and the one by Alizadehsani et al. [9] resorted to ensemble learning and shallow neural networks applied to clinical datasets such as Z-Alizadeh Sani and Cleveland, with modest success in disease classification. However, a sufficient number of problems persist as obstacles to be solved even with these advances.

Firstly, many previous investigations were looking for arrhythmia detection or CAD prediction in deployment and missed the opportunity for integrated cardiac screening matching real-world diagnostic concepts where patients may arrive with co-occurring conditions. Secondly, these models are usually overfitted due to the limited number and imbalance of medical datasets, particularly in clinical datasets like Z-Alizadeh Sani. On the other hand, these CNN-only models would weakly express the temporal patterns in ECG signals, lacking some memory mechanisms to model such sequential dependencies. Third, only a few studies have attempted to establish a systematic AI approach that unites time-series biosignals with structured clinical data for diagnosis in one architecture.

In other words, we propose an AI-based dual-model framework for the complete cardiac disease detection process. More particularly, we utilize a 1D CNN and a hybrid CNN-LSTM for arrhythmia classification using the MIT-BIH Arrhythmia dataset. The CNN extracts local features from the ECG signals, whereas the LSTM captures temporal dependencies, augmenting robustness and accuracy in modeling sequences. For the CAD prediction, a feedforward DNN is built on the Z-Alizadeh Sani dataset. Batch normalization and dropout layers are introduced in the model to ensure generalization and avoid overfitting, leading to stable performance on fairly small datasets.

Figure 1 shows the differences between a normal heart and one that has a VSD. The condition is characterized by an abnormal hole in the wall separating the two ventricles, allowing blood to flow from the left ventricle to the right. This increases the strain on the heart and causes cardiovascular problems. Early diagnosis is crucial for proper treatment and improved results.

1.png

Figure 1. Difference between a normal heart and a heart with a ventricular septal defect, which is a defect at birth (congenital heart defect) [10]

Key features:

(1). CNN-LSTM hybrid network architecture: Utilizes local spatial features along with long-term temporal dependencies extracted from ECG signals for effective detection of arrhythmia.

(2). Regularized DNN: Detects the intricate nonlinear connections between the structured clinical data to avoid overfitting.

(3). Dual-model approach: Combines the use of customized deep learning models for detecting the two different conditions, namely arrhythmia and CAD.

(4). Automatic feature extraction: Removes the need for manual feature engineering by extracting informative features from ECG signals and clinical features automatically.

(5). Comprehensive diagnostic performance comparison: Illustrates enhanced diagnostic performance by comparing the proposed approach with traditional machine learning and deep learning methods.

(6). Applicability in clinical settings: Enables early detection of heart diseases, which can later be used as an input to AI-based decision support systems in clinics.

(7). Generalization ability: Uses regularization strategies to enhance the generalization ability of the model to handle unseen patient data.

(8). Scalable model structure: Offers a scalable architecture that is useful for multimodal cardiovascular disease diagnosis and future intelligent healthcare applications.

Major contributions of this research include providing a framework for diagnosing cardiovascular conditions by integrating deep learning techniques with time series ECG data and structured clinical information. A novel CNN-LSTM network has been designed to enhance the capability of arrhythmia classification by using temporal memory, which is an improvement compared to existing CNNs. Furthermore, a fully connected DNN model predicts CAD on the basis of clinical variables, wherein the use of batch normalization and dropout avoids overfitting. Last but not least, benchmarking datasets, known as MIT-BIH Arrhythmia and Z-Alizadeh Sani CAD, are included in order to test the proposed method in real-world scenarios.

MIT-BIH Arrhythmia dataset is a well-known, annotated ECG dataset made up of 48 recordings of half an hour each from 47 patients recorded at the Beth Israel Hospital Arrhythmia Laboratory [11]. This benchmarking dataset offers a large number of labeled ECG waveform cases belonging to various types of arrhythmias, hence suitable for training deep learning models in time-series classification problems. As opposed to this, the Z-Alizadeh Sani dataset comprises 303 patient data points with 54 different attributes such as demographics, symptoms, ECG, and lab data.

The organization of this research paper is described below. Section 2 describes the relevant literature in the area, highlighting past studies in applying deep learning for ECG signal classification and CAD prediction. Section 3 describes the proposed architecture of dual models that includes the selection of CNN, CNN-LSTM, and DNN architectures. Section 4 gives an overview of experiments performed, describing in detail the process of data preprocessing, model configuration, training, and evaluation, and provides a discussion of the results and their applicability in the clinical setting. Section 5 concludes the whole study.

Being able to detect cardiac abnormalities at the earliest possible stage, especially arrhythmias and CAD, still remains an acute challenge in the healthcare domain, in spite of the astounding pioneering developments in AI and deep learning. Clinical misdiagnosis and the lack of expert interpretations, coupled with conventional manual diagnosis, lead to delays in the administration of therapies and poor patient outcomes. Current AI-based methods either remain isolated from ECG-based arrhythmia detection or clinical feature-based CAD prediction, thus failing to address the multi-condition diagnostic scenario present in real-world cardiology. In addition, many, if not most, of such approaches depend heavily on hand-engineered features with complicated multi-stage pipelines and a potential lack of explicit regularization strategies, thereby compromising their ability to generalize well on unseen patient data.

The core problem addressed in this study is the lack of a unified, end-to-end AI framework capable of accurately detecting multiple cardiac abnormalities using modality-specific deep learning architectures that are optimized for both time-series signals (ECG) and structured clinical datasets. To overcome the identified limitations and bridge the research gap, this study proposes an AI-driven dual-model architecture that combines two tailored deep learning pipelines:

•A 1D CNN and CNN-LSTM hybrid model for arrhythmia detection, using ECG signals from the MIT-BIH Arrhythmia dataset

•A feedforward DNN with dropout and batch normalization for CAD prediction, trained on the Z-Alizadeh Sani clinical dataset.

This bifurcated design ensures that each cardiac condition is addressed using a deep learning model that is best suited to the data modality and underlying pathophysiological patterns.

3.1 Arrhythmia detection using 1D Convolutional Neural Network and Convolutional Neural Network-Long Short-Term Memory

The MIT-BIH Arrhythmia dataset consists of long-term ambulatory ECG recordings, annotated with beat-level classifications. These one-dimensional time-series signals contain rich spatial and temporal information about the electrical behavior of the heart.

3.1.1 Convolutional Neural Network model

According to Algorithm 1, the 1D CNN model is designed to extract local features from ECG sequences, such as P-QRS-T wave shapes and morphological deviations. Multiple convolutional and max-pooling layers are stacked to capture hierarchical representations of the waveform data. The network is followed by fully connected layers for final classification into arrhythmia classes.

Considering $X=\left\{x_1, x_2, \ldots, x_T\right\} \in \Re^{T \times 1}$ be a 1D ECG signal segment of length T, representing a sequence of ECG measurements, a 1D CNN applies convolutional filters to extract spatial (local) features:

$\begin{gathered}h_t^{(1)}=\sigma\left(\sum_{i=0}^{k-1} w_i^{(1)} x_{t+i}+b^{(1)}\right), \\ t=1,2, \ldots, T-k+1\end{gathered}$              (1)

where, $w_i^{(1)}$ are the learnable weights of the convolution filter, $b^{(1)}$ is the bias term, k is the kernel size, σ is the activation function. Typically ReLU(z) = max(0,z).

The output is passed through multiple such layers, often followed by max pooling:

$p_t=\max _{j \in w} h_j^{(1)}$            (2)

where, W is the pooling window.

3.1.2 CNN-LSTM hybrid model

To enhance the model’s capacity to learn long-range temporal dependencies, the CNN is integrated with an LSTM layer. This is shown in Algorithm 2 CNN-LSTM stepwise. The CNN component handles spatial encoding, while the LSTM captures the sequential progression and temporal correlation between heartbeats. This hybrid architecture improves sensitivity to rhythm-based conditions, such as premature ventricular contractions or bundle branch blocks.

Model configuration

•Input: segmented ECG beats (normalized)

•CNN layers: 3–4 convolutional layers with ReLU activation

•LSTM layer: single or stacked layer with 64–128 units

•Dense layers: fully connected layers with Softmax for multi-class classification

•Optimizer: Adam

•Loss function: categorical cross-entropy

•Evaluation: accuracy, precision, recall, F1-score.

The CNN output $H=\left\{h_1, h_2, \ldots, h_T\right\}$ is fed into an LSTM layer to capture temporal dependencies.

For each timestep t, the LSTM equations are:

$f_t=\sigma\left(W_f h_t+U_f h_{t-1}+b_f\right)$             (3)

$i_t=\sigma\left(W_i h_t+U_i h_{t-1}+b_i\right)$            (4)

$\tilde{c}_t=\tanh \left(W_c h_t+U_c h_{t-1}+b_c\right)$               (5)

$c_t=f_t \circ c_{t-1}+i_t \circ \tilde{c}_t$             (6)

$o_t=\sigma\left(W_o h_t+U_o h_{t-1}+b_o\right)$         (7)

$h_t=o_t \circ \tanh \left(c_t\right)$           (8)

$f_t=\sigma\left(W_f h_t+U_f h_{t-1}+b_f\right)$             (9)

where,

$f_t, i_t, o_t$: forget, input, and output gates, respectively

$\circ$: element-wise multiplication

c_t: cell state

h_t: hidden state

$W_*, U_*, b_*$: LSTM parameters

The final hidden state h_T is passed through a fully connected layer and softmax activation for multi-class classification:

$\hat{y}=\operatorname{softmax}\left(W \rightleftarrows_o h_T+b_o\right)$            (10)

where,

$\operatorname{softmax}\left(z_i\right)=\frac{e^{z_i}}{\sum_{j=1}^C e^{z_j}}, i=1, \ldots, C$

C: number of arrhythmia classes.

The model is trained by minimizing categorical cross-entropy:

$L_{\text {arrhythmi }}=-\sum_{i=1}^C y_i \log \left(\hat{y}_i\right)$           (11)

where, $y \in\{0,1\}^C$ is the one-hot encoded ground truth.

3.2 Coronary artery disease prediction using regularized feedforward Deep Neural Network

The Z-Alizadeh Sani dataset includes 303 patient records with 54 features covering demographic, symptom, ECG, echocardiographic, and laboratory variables. This structured dataset is well-suited for classification tasks using deep feedforward networks.

3.2.1 Deep Neural Network architecture

A multi-layer feedforward DNN, which is described in Algorithm 3, is employed to model complex nonlinear relationships among clinical features that contribute to CAD. The network includes several dense layers interleaved with dropout layers for regularization and batch normalization for training stability.

For each hidden layer l, the transformation is:

$a^{(l)}=B N^{(l)}\left(W^{(l)} h^{(l-1)}+b^{(l)}\right), h^{(l)}=\sigma\left(a^{(l)}\right)$            (12)

where,

$W^{(l)} \in \Re^{n_l \times n_{l-1}}$: weight matrix

$B N^{(l)}$: batch normalization

σ: activation function (ReLU)

Dropout is applied as:

$h_{\text {drop }}^{(l)}=h^{(l)} \cdot \delta, \delta \sim$ Bernoulli $(1-p)$

where, p is the dropout rate.

3.2.3 Feature preprocessing

Before training, categorical variables are one-hot encoded, and continuous variables are normalized. Feature importance analysis is optionally conducted to reduce dimensionality and enhance interpretability.

Model configuration

•Input: 54-dimensional feature vector

•Hidden layers: 3–5 dense layers with 64–128 neurons

•Activation: ReLU

•Dropout: 0.3–0.5 after each layer

•Batch normalization: after each hidden layer

•Output: Sigmoid-activated neuron for binary classification

•Optimizer: Adam

•Loss function: binary cross-entropy

•Evaluation: accuracy, area under the curve (AUC)-receiver operating characteristic (ROC), precision, recall, F1-score

Final prediction is made using sigmoid activation:

$\hat{y}=\sigma\left(W^{(L)} h^{(L-1)}+b^{(L)}\right)$              (13)

$\sigma(z)=\frac{1}{1+e^{-z}}$            (14)

where, $y \in\{0,1\}$ represents the probability of CAD presence.

Binary cross-entropy is used as the loss function:

$L_{C A D}=-[y \log (\hat{y})+(1-y) \log (1-\hat{y})]$              (15)

Normalization through batches keeps the feature distribution stable and speeds up convergence because of reduced internal covariate shift. Dropout works by randomly turning off some neurons during training in order to avoid co-adaptation and reduce overfitting. According to ablation experiments, the combination of batch normalization and dropout enhances generalization performance is shown in Table 2.

Table 2. Performance of the combination of batch normalization and dropout

3.3 Unified diagnostic framework

The two models operate in parallel pipelines, each specialized for a specific cardiac abnormality. This architecture offers the following advantages:

•Modality-specific learning: Optimizes performance by using CNN-based models for time-series data and DNNs for clinical feature-based data.

•Multi-condition capability: Enables the simultaneous detection of both arrhythmia and CAD, supporting more comprehensive diagnostic decision-making.

•End-to-end learning: Both models are trained and deployed without reliance on manual feature engineering, ensuring scalability and real-time applicability.

•Regularization and stability: Batch normalization and dropout techniques enhance robustness and reduce overfitting, particularly important for the smaller CAD dataset.

The models are optimized separately using the Adam optimizer to minimize:

$L_{\text {Total }}=L_{\text {arrhythmia }}+L_{C A D}$            (16)

Though the models are independent, this dual-path structure can be deployed in parallel pipelines or integrated within a larger multi-modal framework.

3.4 Feature importance analysis

From feature importance analysis, it was evident that RR interval, QRS interval, and QT interval were the most discriminating features from an ECG perspective for classification of arrhythmias. Chest pain type, age, cholesterol level, blood pressure, and peak heart rate were the most influential clinical factors when predicting coronary artery disease (see Table 3). The combination of these factors made the proposed approach successful in cardiac abnormality detection.

Table 3. Important features contributing to cardiac abnormality detection

3.5 Model selection

The choice of various deep learning structures depended on the nature of the input data and the requirement of diagnosis in each case. The process of diagnosing arrhythmia depends on the utilization of ECG signal data that is sequential in nature, having both local morphological patterns as well as temporal relationships. Hence, the selection of the CNN-LSTM structure was considered optimal since CNNs are able to capture local features such as P-waves, QRS complex, and T-waves while LSTMs are utilized to capture temporal dependencies between successive heartbeats. It has been proven in past studies that hybrid CNN-LSTM models outperform CNN and LSTM models alone due to joint learning of spatial and temporal representation [7, 15, 18].

However, in the case of CAD diagnosis in this experiment, the utilization of the Z-Alizadeh Sani dataset takes place, where structured clinical variables are involved, which include demographics, laboratory data, symptoms, and risk factors, and does not involve any ECG signals, which are sequential in nature and do not include temporal dependencies. Hence, a recurrent structure like LSTM is not required in this case. Hence, the use of a DNN was considered to be optimal since it learns the complex relationship between structured clinical data. The inclusion of dropout and batch normalization further improves generalization and reduces overfitting, making the DNN well suited for structured clinical data [9, 17, 19].

Alizadehsani et al. [30] developed a data mining approach for coronary artery disease diagnosis using clinical and laboratory features. However, the study did not incorporate ECG signals or multimodal learning, limiting its ability to capture complex cardiac patterns.

This choice of architecture is not arbitrary but data-driven, meaning that a specific architecture will be matched to the type of input data. The choice of such an approach allows efficient learning of features from ECG signals and clinical data, keeping computational efficiency intact.

The architecture makes it possible for the proposed framework to leverage the advantages of both DL models and thus provide a better approach to detect any heart abnormality compared to the scenario where one model performs both tasks.

The proposed study offers a combination of hybrid deep learning approaches consisting of a CNN-LSTM network to detect arrhythmias based on ECGs and a regularized DNN to predict CAD based on structured clinical data. As shown in experimental studies with the MIT-BIH and Z-Alizadeh Sani datasets, the proposed framework is superior to traditional machine learning methods as well as deep learning techniques in terms of accuracy, precision, recall, F1-score, and AUC-ROC metrics. The combination of CNN-LSTM and DNN allows for efficient learning from diverse cardiac data.

However, the study suffers from a number of drawbacks. Firstly, the suggested framework is tested only on publicly available benchmarks with limited data diversity, which does not provide an accurate picture of real-world patients. Secondly, the arrhythmia and CAD models are trained separately on the corresponding datasets, but not as one multimodal dataset, which means that there is no end-to-end learning. Thirdly, external validation and prospective studies have not been carried out, and the performance of the model in other hospitals, among different demographic groups, and on various data acquisition devices needs to be studied.

The future work will involve designing a comprehensive multimodal framework that can learn from both ECG data and structured clinical data within one architecture. The implementation of mechanisms such as attention, graph neural networks, or transformers will be considered in order to enhance feature extraction and model interpretability. Furthermore, future work will incorporate external multi-center validation using large datasets, explainable AI techniques for better acceptance by clinicians, and model optimization for faster processing.