A Hybrid Decision Tree and Attention-Based Neural Network (DT-ATTN) Framework for Arrhythmia Detection and Multi-Class Classification

Abstract

Cardiac arrhythmias are serious health hazards that need precise, understandable diagnostic instruments. This study proposes an Explainable Hybrid DT-ATTN Model that combines Decision Trees (DT) and an Attention-Based Neural Network (ATTN) to detect and classify arrhythmias using real-time Electrocardiogram (ECG) data. The Biocare ECG-1210 equipment was used to gather a dataset of 30,000 ECG signals from the Sher-i-Kashmir Institute of Medical Sciences (SKIMS), Soura, Srinagar. The Pan-Tompkins technique and Min-Max Normalisation were used for data pre-processing. The model was trained and validated using an augmented dataset of 72,000 ECG records that included five classes: normal sinus rhythm, Atrial Fibrillation, Ventricular Tachycardia, bradycardia, and Atrial Flutter. While the attention technique enhances feature extraction and enables the model to focus on key ECG patterns, the decision tree component ensures interpretability. In every class, the Hybrid DT-ATTN Model achieves an accuracy of 97%, a precision of 97.01%, a recall of 97.03%, and an F1-score of 97.39% demonstrating its outstanding performance. Additionally, SHAP and LIME were used to illustrate the model's explainability, providing doctors with clear insights into how it makes decisions. The suggested DT-ATTN framework is a dependable tool for real-time arrhythmia diagnosis, as it not only provides high diagnostic accuracy but also bridges the gap between clinical interpretability and the complexity of deep learning models. This study demonstrates how accurate and comprehensible ECG analysis, utilising hybrid AI models, can improve patient outcomes.

Keywords

cardiac arrhythmias electrocardiogram decision tree attention-Based neural network classify

1 Introduction

Cardiovascular diseases (CVDs) are the leading cause of mortality worldwide, accounting for approximately 17.9 million deaths each year, which represents about 32% of all global deaths (World Health Organisation). Among these, cardiac arrhythmias remain a significant concern due to their high prevalence and clinical severity. Atrial fibrillation (AFib), the most common sustained arrhythmia, affects more than 33 million people globally, and its prevalence is expected to double by 2050 as populations continue to age (The Lancet). Ventricular arrhythmias, on the other hand, are a primary cause of sudden cardiac death, which contributes to 15–20% of all deaths worldwide annually (Circulation, American Heart Association). These statistics underscore the pressing need for reliable and effective arrhythmia detection frameworks to facilitate timely diagnosis and enhance patient outcomes.

Cardiac Arrhythmias are Irregular heart rhythms that differ from the normal sinus rhythm, including AFib, ventricular tachycardia (VT), bradycardia, and atrial flutter (shown in Figure 1) (Katal et al., 2023). These anomalies might be sporadic or chronic, which makes it difficult to identify and categorise them (Sheerinsithara and Raj, 2025). AFib, bradycardia, VT, and atrial flutter are examples of irregular heart rhythms that deviate from the typical sinus rhythm (Han et al., 2024). It is challenging to recognise and classify these anomalies since they may be intermittent or persistent. Explainable AI (XAI) is a field of artificial intelligence that focuses on creating models that offer clear and understandable insights into how they make decisions (Daydulo et al., 2023). Explainability is essential in the healthcare industry for fostering clinician trust and guaranteeing the secure implementation of AI-powered solutions.

Figure 1.

Normal sinus rhythm and irregular heart rhythms.

Recent advances in machine learning (ML) and deep learning (DL) have shown great promise in the automated identification of arrhythmias. Traditional approaches, such as rule-based systems and traditional ML models (e.g., support vector machines and random forests), have been widely employed; however, they often lack the accuracy and stability necessary for clinical applications (Daydulo et al., 2023; Ismail et al., 2023; Katal et al., 2023). DL models, notably convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have achieved cutting-edge performance in ECG analysis due to their ability to automatically extract complex features from raw data (Han et al., 2024; Sheerinsithara and Raj, 2025). However, these models are sometimes criticised for their “black-box” nature, which limits their interpretability and prevents their use in clinical practice.

To overcome this limitation, hybrid models that integrate interpretable components with DL techniques have been proposed (Daydulo et al., 2023). These models aim to strike a balance between accuracy and explainability, making them more suitable for healthcare applications (Ismail et al., 2023). Despite their potential, research in this field is restricted, with few papers investigating the integration of attention mechanisms—a powerful DL technique for feature extraction—with interpretable models such as decision trees (DT). Furthermore, most recent research utilises benchmark datasets, such as MIT-BIH, which may not accurately represent the variability and complexity of real-world ECG data.

The goal of this research is to create an Explainable Hybrid DT-ATTN Model that combines DT and an attention-based neural network (ATTN) to detect arrhythmias accurately and interpretably using real-time ECG data. The proposed model aims to address the limitations of existing approaches by leveraging the interpretability of DT and the attention mechanism's feature extraction capabilities, resulting in a trustworthy tool for real-time arrhythmia diagnosis.

The structure of the research paper is as follows: A detailed review of existing studies and related literature is presented in the Literature Review (Section 2). The research methodology, which includes data collection, data pre-processing, feature extraction, and classification, is described in the Proposed Methodology (Section 3). The experimental results, performance analysis, and comparisons with current models are presented in the Results and Discussion (Section 4). The limitations and possible enhancements for further research are emphasised in the Challenges and Future Directions (Section 5). The study's main conclusions and contributions are outlined in the Conclusion (Section 6). Lastly, a list of all the sources cited in the paper can be found in the References (Section 7).

2 Literature Review

The automated detection of CVDs, particularly cardiac arrhythmias, has progressed significantly with the adoption of ML and DL techniques. Improvements in computational efficiency, the availability of large-scale ECG datasets, and the emergence of XAI have collectively enhanced the reliability and clinical relevance of ECG-based diagnostic systems. This section briefly reviews key developments while highlighting persistent research gaps.

Early comparative investigations demonstrated the superiority of DL models over traditional ML techniques. Katal et al. (2023) showed that CNN- and RNN-based architectures achieved higher arrhythmia classification accuracy because they can automatically learn discriminative features from ECG signals. However, the absence of external validation limited the clinical generalizability of their findings. Building on this, Sheerinsithara and Raj (2025) proposed a hybrid ML–DLNN framework that combined handcrafted and learned features, improving classification performance. Despite these gains, the study relied on a small dataset and lacked patient-wise validation, raising concerns regarding robustness.

To address deployment challenges, Han et al. (2024) introduced a transformation-based DL model integrated with edge computing, enabling low-latency, real-time arrhythmia detection. While effective for scalable deployment, the model did not evaluate long-term stability or performance under noisy ECG conditions. Similarly, Daydulo et al. (2023) employed time–frequency representations to capture nonlinear ECG characteristics, thereby improving feature extraction. Nevertheless, increased computational complexity and limited multi-dataset evaluation constrained broader applicability.

Reinforcement learning (RL) has also been explored to enhance adaptability. Ismail et al. (2023) proposed RL-ECGNet, a resource-aware framework capable of multi-class arrhythmia detection under constrained computational settings. Although suitable for embedded systems, the approach lacked interpretability analysis and comprehensive baseline comparisons.

Beyond core DL architectures, several studies have contributed complementary perspectives. Feature selection strategies were investigated to improve ML-based CVD detection efficiency (Ullah et al., 2024), while disease-specific frameworks targeted specialised populations such as diabetic patients (Mokesh Rayalu et al., 2024). Traditional ML approaches remain clinically relevant when paired with effective pre-processing (Jat et al., 2024). Edge-AI solutions have further extended arrhythmia detection to resource-constrained and extreme environments (Mani et al., 2024). Hybrid deep architectures (Denysyuk et al., 2023; Issa et al., 2023; Talukder et al., 2024), IoT-enabled ECG systems (Sai Kumar et al., 2023), ensemble learning methods (Ramkumar et al., 2023), and optimised single-lead ECG models (Karthiga and Santhi, 2022; Madan et al., 2022) have expanded applicability to wearable and remote healthcare scenarios. Earlier inter- and intra-patient modelling approaches also laid important foundations for generalisation-aware learning (Mousavi and Afghah, 2019).

The availability of large, well-annotated datasets has been instrumental in advancing this field. The 12-lead ECG database introduced by Zheng et al. (2020) significantly improved benchmarking and generalisation, while broader clinical perspectives on AI in electrophysiology contextualised algorithmic progress within real-world practice (Feeny et al., 2020). Earlier ML-based rhythm detection studies (Figuera et al., 2016; Ganesh Kumar and Kumaraswamy, 2014; Mitra and Samanta, 2013) established baseline methodologies that continue to influence modern designs.

More recent contributions include automated ML-based arrhythmia-detection systems (Tutuko et al., 2021), multi-band nonlinear ML frameworks for cardiovascular diagnosis (Ribeiro et al., 2024), and heartbeat-feature-driven ML classifiers that bridge arrhythmia detection with broader CVD assessment (Alarsan and Younes, 2019). Hambarde et al. (2025) demonstrated effective arrhythmia detection using a hybrid deep neural network on single-lead ECG data, highlighting suitability for wearable devices.

Advanced architectures integrating transformers have also emerged. Kim et al. (2025) proposed a hybrid CNN-Transformer model using the Stockwell transform, eliminating the need for R-peak detection and enabling end-to-end classification. However, the computational cost and reduced interpretability of transformer-based models remain concerns. Lamba et al. (2025) introduced the FADLEC framework, integrating feature extraction and classification within a unified DL model, though robustness under noise and cross-device variability was not evaluated.

Addressing efficiency, Elsheikhy et al. (2025) proposed a lightweight DL architecture for multi-lead ECG analysis suitable for constrained environments, but real-time deployment on actual edge devices was not validated. Finally, comprehensive reviews by Reshad et al. (2025) synthesised recent advances in DL-based arrhythmia detection, outlining both technical progress and clinical challenges, while underscoring the need for experimentally validated, interpretable, and computationally efficient solutions.

To synthesise the critical analysis presented in the preceding narrative review, Table 1 provides a structured summary of the state of the art in ECG-based arrhythmia detection. This tabular overview crystallises the dominant trends and, more importantly, the persistent methodological gaps identified across the literature. It systematically juxtaposes the technical approaches and their reported performance against fundamental limitations in generalizability, computational efficiency, and clinical practicality. The consolidation of this information objectively delineates the research landscape, offering a definitive reference that underscores the necessity of the novel contribution proposed in this work.

Table 1.
Summary of ECG Arrhythmia Detection Methods, Highlighting the Critical Gap Between High Benchmark Accuracy and Limitations in Clinical Deployment.

Author(s) & Year Core Technique/Model Dataset (Size) Problem Focus/Contribution Reported Performance (%) Identified Research Gap/Limitation

Katal et al. (2023) 1D-CNN, LSTM, GRU MIT-BIH (48 records) 5-class arrhythmia comparison 99.20 Single dataset; no novelty; class imbalance

Sheerinsithara and Raj (2025) ML–DLNN MIT-BIH Hybrid heartbeat detection 99.87 No real-time analysis; single dataset

Han et al. (2024) Transformation DL MIT-BIH Edge-optimized ECG classification 98.84 Limited edge/noisy evaluation

Daydulo et al. (2023) CNN + CWT MIT-BIH (∼109k beats) Time–frequency arrhythmia detection 98.71 High computation; single dataset

Ismail et al. (2023) RL-ECGNet MIT-BIH Resource-aware arrhythmia detection 98.70 Complex RL; limited interpretability

Ullah et al. (2024) ML + Feature Selection UCI (303) CVD detection 98.20 Small dataset; not ECG-centric

Mokesh Rayalu et al. (2024) ADAPT Framework Private (Diabetic) Disease-specific arrhythmia detection 99.12 Limited generalizability

Jat et al. (2024) Traditional ML PhysioNet Arrhythmia detection 97.30 Low novelty; vague method

Mani et al. (2024) ML Workflow PhysioNet Edge arrhythmia detection 96.00 Lower accuracy; niche use

Talukder et al. (2024) ConvNeXt-X MIT-BIH High-accuracy classification 99.90 Very high computation

Issa et al. (2023) 1D-CNN MIT-BIH (Lead II) Single-lead classification 99.50 Limited ECG information

Sai Kumar et al. (2023) ResNet + IoT MIT-BIH IoT-based detection 99.35 System-focused contribution

Ramkumar et al. (2023) Ensemble ML MIT-BIH Ensemble arrhythmia detection 99.73 High latency

Madan et al., (2022) CNN–LSTM MIT-BIH Spatio-temporal ECG learning 99.21 Black-box; costly

Mousavi and Afghah (2019) Seq-to-Seq DL MIT-BIH Patient-aware classification 99.00 High complexity

Kim et al. (2025) CNN–Transformer MIT-BIH No R-peak detection 99.10 High preprocessing cost

Elsheikhy et al. (2025) Lightweight DL CPSC 2018 Efficient multi-lead ECG 99.60 Hardware dependent

Author(s) & Year	Core Technique/Model	Dataset (Size)	Problem Focus/Contribution	Reported Performance (%)	Identified Research Gap/Limitation
Katal et al. (2023)	1D-CNN, LSTM, GRU	MIT-BIH (48 records)	5-class arrhythmia comparison	99.20	Single dataset; no novelty; class imbalance
Sheerinsithara and Raj (2025)	ML–DLNN	MIT-BIH	Hybrid heartbeat detection	99.87	No real-time analysis; single dataset
Han et al. (2024)	Transformation DL	MIT-BIH	Edge-optimized ECG classification	98.84	Limited edge/noisy evaluation
Daydulo et al. (2023)	CNN + CWT	MIT-BIH (∼109k beats)	Time–frequency arrhythmia detection	98.71	High computation; single dataset
Ismail et al. (2023)	RL-ECGNet	MIT-BIH	Resource-aware arrhythmia detection	98.70	Complex RL; limited interpretability
Ullah et al. (2024)	ML + Feature Selection	UCI (303)	CVD detection	98.20	Small dataset; not ECG-centric
Mokesh Rayalu et al. (2024)	ADAPT Framework	Private (Diabetic)	Disease-specific arrhythmia detection	99.12	Limited generalizability
Jat et al. (2024)	Traditional ML	PhysioNet	Arrhythmia detection	97.30	Low novelty; vague method
Mani et al. (2024)	ML Workflow	PhysioNet	Edge arrhythmia detection	96.00	Lower accuracy; niche use
Talukder et al. (2024)	ConvNeXt-X	MIT-BIH	High-accuracy classification	99.90	Very high computation
Issa et al. (2023)	1D-CNN	MIT-BIH (Lead II)	Single-lead classification	99.50	Limited ECG information
Sai Kumar et al. (2023)	ResNet + IoT	MIT-BIH	IoT-based detection	99.35	System-focused contribution
Ramkumar et al. (2023)	Ensemble ML	MIT-BIH	Ensemble arrhythmia detection	99.73	High latency
Madan et al., (2022)	CNN–LSTM	MIT-BIH	Spatio-temporal ECG learning	99.21	Black-box; costly
Mousavi and Afghah (2019)	Seq-to-Seq DL	MIT-BIH	Patient-aware classification	99.00	High complexity
Kim et al. (2025)	CNN–Transformer	MIT-BIH	No R-peak detection	99.10	High preprocessing cost
Elsheikhy et al. (2025)	Lightweight DL	CPSC 2018	Efficient multi-lead ECG	99.60	Hardware dependent

Despite advances in automated arrhythmia identification, significant gaps persist. Hybrid models that combine interpretable DT with DL remain underexplored, restricting clinical application due to a lack of transparency. Most studies use benchmark datasets like MIT-BIH, which do not account for real-world issues like noise and artefacts. It is critical to conduct research on real-time clinical ECG data. Furthermore, XAI remains limited, and issues in recognising unusual or overlapping arrhythmias require further research. Addressing these shortcomings is critical for creating accurate, interpretable, and clinically useful models.

The proposed model directly addresses the scarcity of research on hybrid interpretable models for ECG analysis and the unique challenge of integrating real-time clinical data into AI-driven diagnostic tools. By emphasising explainability and real-world applicability, this work aims to address the limitations of prior research and advance automated arrhythmia identification. Attention mechanisms enhance the model's ability to focus on crucial ECG patterns, while DT provide clear insights into the decision-making process, bridging the gap between complex AI models and clinical interpretability.

3 Materials and Methods

This paper introduces a Hybrid Decision Tree and Attention-Based Neural Network (DT-ATTN) Framework for detecting multi-class arrhythmias using real-time ECG data. The institutional Ethical Committee of Lovely Professional University issued ethical approval for this research (Ref No.: IEC-LPU/2025/1/19). The methodology is divided into three main stages: data collection and pre-processing, feature extraction, and classification. Each stage is critical to establishing accurate, interpretable, and robust arrhythmia categorisation.

3.1 Data Collection

In this work, 30,000 ECG signals from the Sher-i-Kashmir Institute of Medical Sciences (SKIMS), located in Soura, Srinagar, were used as the dataset. The Biocare ECG-1210 equipment was used to capture the ECG signals, which had a fixed sampling rate of 250 Hz and lasted 10 s each. To ensure efficient processing, we assigned numerical labels to each arrhythmia class: 0 for Normal Sinus Rhythm, 1 for AFib, 2 for VT, 3 for Bradycardia, and 4 for Atrial Flutter. Although the dataset was initially balanced, data augmentation using SMOTE was applied to expand the sample size and introduce greater variability in the ECG recordings. This approach enhances the model's ability to generalise, reduces the risk of overfitting, and ensures improved stability and reliability in classifying different arrhythmia types. Graph 1 illustrates the class distribution before data augmentation, with each class comprising approximately 6,000 samples and balanced evenly. Graph 2 illustrates the class distribution after augmentation, where the dataset size has increased to approximately 14,500 samples per class, thereby maintaining a balance across all classes

Graph 1.

Class Distribution of the original dataset [AQ10].

Graph 2.

Class distribution of the augmented dataset.

Figure 2.

Overall methodology.

3.2 Data Pre-Processing

The raw ECG signals underwent several pre-processing procedures to ensure high-quality inputs for model training. The detection of R-peaks, which is essential for precise heartbeat segmentation, was done using the Pan-Tompkins algorithm. To accurately extract features in the following steps, this approach employs bandpass filtering, differentiation, squaring, and moving window integration to locate the R-peaks. ECG data often contain noise from various sources, including electrode contact artefacts and muscle movements. The therapeutically relevant elements of the ECG waveform were preserved while baseline drift and high-frequency noise were eliminated using a bandpass filter (0.5–50 Hz). To ensure consistency across all signals, Min-Max Normalisation was applied, scaling the ECG amplitudes within the range of [−1, 1]. This normalisation method helps maintain numerical stability during model training and improves classification performance. We increased the accuracy and dependability of the model by using these pre-processing approaches to ensure the ECG signals used for training and validation were clear, well-segmented, and normalised.

3.3 Feature Extraction

The pre-processed ECG signals were subjected to a thorough extraction of time-domain, frequency-domain, and morphological features in order to improve classification accuracy. To assess heartbeat fluctuations and identify arrhythmias, such as AFib, time-domain parameters, including RR intervals and QRS duration, were calculated. Fast Fourier Transform (FFT) coefficients and other frequency-domain properties helped distinguish between different types of arrhythmias by offering insights into the spectrum characteristics of ECG data. Furthermore, waveform shape variations were captured by morphological aspects, including P-wave amplitude and T-wave variability, which are crucial for detecting disorders such as atrial flutter. To ensure a well-organised dataset for model training, these extracted characteristics were arranged using Biocare 1000 Data Management Software and saved in a .csv format.

3.4 Classification

The proposed Hybrid DT-ATTN Model combines an ATTN for enhanced feature extraction and classification accuracy with DT for interpretability (shown in Figure 2). The DT component offers a structured, rule-based approach that makes the model's predictions transparent and interpretable for clinicians, while the ATTN mechanism improves classification by emphasising the most relevant signal features, allowing the model to prioritise essential ECG segments. The model was trained on an augmented dataset of 72,000 ECG signals, optimised using the Adam optimiser, and assessed using accuracy, precision, recall, and F1-score, achieving over 97% accuracy across all arrhythmia classes. Explainability was further validated via feature importance analysis and attention weight visualisation, guaranteeing dependable and clinically interpretable decision-making.

The hybrid DT-ATTN model enhances the accuracy and explainability of arrhythmia detection by combining interpretable decision tree classification with a DL-based attention mechanism. A detailed explanation of how the results are combined to yield the final prediction, which is presented below, is provided.

A DT is a hierarchical model that predicts a class label using learned decision rules. Given an ECG set of parameters $X = {x_{1}, x_{2}, \dots x_{n}}$ , the DT model uses recursive binary splitting to select features that minimise the Gini Impurity.

\begin{aligned} G_{i n i} (D) = 1 - \sum_{i = 1}^{c} p_{L}^{2} \end{aligned}

(1)

where p is the probability of class i in dataset D, and C is the total number of classes. The decision tree assigns a probability score for each class:

\begin{aligned} F_{D T} = P (Y | X, θ_{D^{T}}) \end{aligned}

(2)

Where

$F_{D T}$ is the predicted probability distribution from the DT mode l

$X$ represents input ECG features

$θ_{D T}$ represents learned decision rules .

An attention mechanism assigns different importance scores to ECG features before classification. Given input feature vectors X the attention scores are computed as:

\begin{aligned} e_{i} = v^{T} tanh (W x_{l} + b) \end{aligned}

(3)

where:

W is the weight matrix

v is a learnable weight vector

b is a bias term

tanh(·) is the activation function.

The normalised attention scores $α_{i}$ are obtained using the softmax function:

\begin{aligned} α_{i} = \frac{e x p (e_{i})}{\sum_{j = 1}^{n} e x p (e_{j})} \end{aligned}

(4)

The final attention-weighted feature representation is computed as:

\begin{aligned} F_{A T T N} = \sum_{i = 1}^{n} a_{i} x_{i} \end{aligned}

(5)

Where $F_{A T T N}$ is the feature representation obtained using attention.

Finally, the ATTN classifies the ECG signal using a softmax function:

\begin{aligned} P (Y | X, θ_{A T T N}) = s o f t m a x (W_{A T T N} F_{A T T N}) \end{aligned}

(6)

Where:

$W_{A T T N}$ is the weight matrix of the classification laye r

$F_{A T T N}$ represents the parameters of the ATTN.

To combine the outputs of both models, we compute a weighted sum of the probability distributions obtained from DT and ATTN:

\begin{aligned} \hat{Y} = σ (W_{D T} F_{D T} + W_{A T T N} F_{A T T N}) \end{aligned}

(7)

where:

is the final predicted probability distribution over arrhythmia classes

$W_{D T}$ and $W_{A T T N}$ are learnable weight parameters for DT and ATTN output s

σ is the softmax activation function that ensures probabilities sum to 1.

Expanding σ, we get:

\begin{aligned} {\hat{y}}_{i} = \frac{e x p (W_{D T} F_{D T} + W_{A T T N} F_{A T T N})}{\sum_{j = 1}^{C} e x p (W_{D T} F_{D T} + W_{A T T N} F_{A T T N})} \end{aligned}

(8)

for each class i, where C is the number of classes.

The model is optimised by minimising the cross-entropy loss:

\begin{aligned} L = - \sum_{i = 1}^{c} y_{i} l o g ({\hat{y}}_{i}) \end{aligned}

(9)

where:

$y_{i}$ is the true label (one-hot encoded)

${\hat{y}}_{i}$ is the predicted probability for class i .

Table 2 presents the step-by-step pseudocode of the proposed DT-ATTN Hybrid Model for arrhythmia detection. The algorithm demonstrates the sequential processing of ECG features through two complementary branches—DT-ATTN, followed by a fusion mechanism and cross-entropy optimisation to yield the final classification outcome.

Table 2.

Pseudocode Representation of the DT-ATTN Hybrid Model for Arrhythmia Detection.

Step	Description
Input	ECG feature vector ( $X = {x_{1}, x_{2}, \dots x_{n}}$ )
Output	Final class prediction ( $\hat{Y}$ )
1. Decision Tree Branch
1.1	Train Decision Tree (DT) using features ( $X$ )
1.2	For each node, compute Gini Impurity: $G_{i n i} (D) = 1 - \sum_{i = 1}^{c} p_{L}^{2}$
1.3	Select feature split minimizing impurity
1.4	Compute probability distribution: $F_{D T} = P (Y \| X, θ_{D^{T}})$
2. Attention-Based Neural Network Branch
2.1	Initialize weights (W), vector (v), and bias (b)
2.2	For each feature (X), compute attention score: $e_{i} = v^{T} \tanh (W x_{l} + b)$
2.3	Normalize scores with softmax: ( $α_{i} = \frac{\exp (e_{i})}{\sum_{j = 1}^{n} e x p (e_{j})}$ )
2.4	Compute weighted feature representation ( $F_{A T T N} = \sum_{i = 1}^{n} a_{i} x_{i}$ )
2.5	Classify with softmax layer: $P (Y \| X, θ_{A T T N}) = s o f t m a x (W_{A T T N} F_{A T T N})$
3. Fusion Layer
3.1	Combine DT and ATTN outputs: $\hat{Y} = S o f t m a x (W_{D T} F_{D T} + W_{A T T N} F_{A T T N})$
4. Optimisation
4.1	Define cross-entropy loss: ( $L = - \sum_{i = 1}^{c} y_{i} \log ({\hat{y}}_{i})$ )
4.2	Update parameters (W, v, b, $W_{A T T N}$ , $W_{D T}$ ) using Gradient Descent/Adam
5. Prediction Phase
5.1	For new ECG input ( $X$ ): obtain $W_{D T}$ ), ( $F_{A T T N}$ ), and compute ( $\hat{Y}$ )
5.2	Output predicted class: (Predicted_Class = arg max ( $\hat{Y}$ ))

To evaluate the robustness and generalisation capability of the proposed hybrid Decision Tree-Attention (DT-ATTN) model for cardiac arrhythmia classification, we employed a 10-fold cross-validation strategy. In this method, the dataset is divided into ten equal subsets. Each subset serves as a validation set once, while the remaining nine subsets are used for training. We recorded the classification accuracy for each fold and calculated the mean and standard deviation to assess the overall stability and reliability of the model. This approach ensures that the model's performance is evaluated across multiple partitions of the dataset, reducing potential bias and providing a comprehensive assessment of its predictive capability.

4 Results and Discussion

The Hybrid DT-ATTN Model was validated using standard methods and demonstrated remarkable classification performance, with an overall accuracy of 97%, a precision of 97.01%, a recall of 97.03%, and an F1-score of 97.39%. Themodel'sprecision, recall, and F1 score were consistently excellent across all arrhythmia classifications, confirming its reliability for real-time ECG analysis. The developed DT-ATTN model was assessed using standard performance metrics:

\begin{aligned} A c c u r a c y & = \frac{T P + T N}{T P + T N + F P + F N} \times 100 \end{aligned}

(10)

\begin{aligned} P r e c i s i o n & = \frac{T P}{(T P + F P)} \times 100 % \end{aligned}

(11)

\begin{aligned} R e c a l l & = \frac{T P}{T P + F N} \times 100 % \end{aligned}

(12)

\begin{aligned} R e c a l l & = \frac{T P}{T P + F N} \times 100 % \end{aligned}

(13)

Where

TP = True Positives

TN = True Negatives

FP = False Positives

FN = False Negatives.

The 10-fold cross-validation results demonstrated that the DT-ATTN model achieved consistently high accuracy across all folds. The accuracies for the individual folds were 97.12%, 97.44%, 97.49%, 97.10%, 96.96%, 97.25%, 97.03%, 97.29%, 96.93%, and 97.07%, resulting in an average accuracy of 97.17% with a standard deviation of ±0.18%. These results indicate that the proposed hybrid model performs reliably across different subsets of the dataset, confirming its effectiveness for multi-class arrhythmia detection and its potential applicability in real-world clinical settings.

The Receiver Operating Characteristic (ROC) Curve and Area Under Curve (AUC) were used to assess the model's ability to differentiate between different arrhythmia classes. The ROC curve compares the True Positive Rate (TPR) to the False Positive Rate (FPR) at different classification thresholds, displaying the trade-off between sensitivity and specificity.

\begin{aligned} T P R & = \frac{T P}{T P + F N} \end{aligned}

(14)

\begin{aligned} F P R & = \frac{F P}{F P + T N} \end{aligned}

(15)

\begin{aligned} A U C & = \int_{0}^{1} T P R (F P R) d (F P R) \end{aligned}

(16)

The per-class analysis of the ROC curves reveals outstanding classification performance across all categories, as shown in Figure 3. The Area Under the Curve (AUC) values are exceptionally high, with Class 0 and Class 4 achieving a perfect score of 1.00, while Class 2 also performs near-perfectly with an AUC of 0.99. Classes 1 and 3 follow closely with a strong AUC of 0.96. Since all AUC scores are substantially greater than the 0.5 benchmark of a random classifier, this indicates that the model possesses a robust and consistent ability to distinguish each individual class from all others, with minimal misclassification. The results confirm the model's high discriminatory power and reliability across the entire classification scheme.

Figure 3.

ROC curve of all classes.

The confusion matrix (as shown in Figure 4) demonstrates the strong performance of the proposed model in classifying different arrhythmia types, with the majority of predictions aligned along the diagonal. Classes 0 and 4 were perfectly classified, and Classes 1, 2, and 3 also had high correct classification rates with only a littleerlap. This result highlights the robustness and reliability of the model, confirming its effectiveness in distinguishing between complex ECG patterns and supporting its potential for accurate clinical decision-making in arrhythmia detection.

Figure 4.

Confusion matrix for the DT-ATTN model on the five-class arrhythmia classification task, showing high accuracy along the diagonal.

SHAP interaction values (as shown in Figure 5) offer a detailed view of how pairs of features jointly influence a model's predictions, providing insight beyond individual feature contributions. In this analysis, interactions were evaluated across five key cardiac features: PR interval, QRS duration, heart rate (HR), systolic blood pressure (BP), and QT interval. Each point in the interaction plot represents a sample, with the SHAP interaction value quantifying the combined effect of a feature pair on the prediction. The colour gradient indicates the relative value of the feature on the y-axis. Notably, significant interactions were observed between the HR and QT intervals, suggesting a physiologically relevant relationship that may influence predictive performance. These findings emphasise the importance of accounting for feature interactions in clinical model interpretation.

Figure 5.

SHAP summary plot displaying feature importance and impact direction for the proposed model, highlighting heart rate and QT interval as the most influential predictors in arrhythmia classification.

The LIME (Local Interpretable Model-Agnostic Explanations) graph for class 1 (AFib) visually highlights the most influential ECG and cardiac features driving the model's prediction. In this explanation, features like HR, diastolic BP, and QT interval positively affect the prediction, meaning that higher values of these measurements make it more likely for the arrhythmia to be classified as AFib. Conversely, features such as systolic BP and T axis contribute negatively, slightly reducing the predicted probability. This interpretability analysis enables clinicians to understand which physiological measurements most impact the model's decisions, enhancing trust in automated arrhythmia detection and offering potential insights for personalised patient monitoring and intervention strategies. Figure 6 shows the LIME graph for class 1, illustrating the relative contributions of each feature to the prediction.

Figure 6.

LIME explanation for class 1 (atrial fibrillation) showing the contribution of key ECG and cardiac features.

The proposed Hybrid DT-ATTN framework outperforms both DL and conventional approaches in key assessment criteria for arrhythmia detection and multi-class classification (as shown in Figure 7). DT-ATTN with an accuracy of 97%, a precision of 97.01%, a recall of 97.03%, and an F1-score of 97.39% when compared to methods such as DT, SVM, Random Forest, CNN, RNN, and LSTM, as demonstrated in Table 3. The model's capacity to recognise intricate ECG patterns is enhanced by combining decision tree-based feature selection with attention-based DL, making it the most efficient technique for classifying arrhythmias.

Figure 7.

Performance comparison of arrhythmia detection methods with DT-ATTN.

Table 3.

Comparison of Different Methods with the Proposed Method (DT-ATTN).

Method	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	AUC-ROC (%)
Traditional Decision Tree (DT)	85.6	83.9	84.3	84.1	88.2
Support Vector Machine (SVM)	89.2	87.8	88.1	88	90.5
Random Forest (RF)	91.4	90.1	90.6	90.3	92.8
K-Nearest Neighbours (KNN)	87.8	85.9	86.2	86	89.1
Convolutional Neural Network (CNN)	92.5	91.7	91.2	91.4	94.3
Recurrent Neural Network (RNN)	93.2	92.5	92.1	92.3	94.9
Long Short-Term Memory (LSTM)	94.1	93.5	93.2	93.3	95.2
Transformer-Based Model	95.4	90.9	95.1	95	96.7
Hybrid CNN-LSTM	95.9	91.3	94.6	95.4	97.1
DT-ATTN (Proposed Method)	97.0	97.0	97.0	97.0	97.3

5 Limitations and Future Directions

While the proposed DT-ATTN framework demonstrates excellent performance, this study has certain limitations that warrant acknowledgement. The model was developed and validated using data from a single medical institution, which may affect its generalizability across diverse populations and clinical settings. Additionally, the evaluation focused on five specific arrhythmia types, excluding other clinically significant rhythm abnormalities. The computational complexity of the hybrid architecture, though providing valuable interpretability, may present challenges for deployment on resource-constrained devices without further optimisation. Future research should focus on multi-centre validation studies, expansion to include a broader spectrum of arrhythmias, and development of optimised versions for wearable device implementation to enhance clinical applicability and adoption.

6 Conclusion

This study introduces a novel Hybrid DT-ATTN model for accurate and interpretable detection of cardiac arrhythmias using real-time ECG data. Trained on an augmented dataset of 72,000 ECG signals, the model achieved consistently high performance, with an accuracy of 97%, a precision of 97.01%, a recall of 97.03%, and an F1-score of 97.39%, as validated through 10-fold cross-validation. The ROC–AUC values, approaching 1.0 across all classes, further confirm its strong discriminatory capability. Beyond performance, the framework ensures clinical interpretability by integrating feature importance analysis, SHAP interaction values, and LIME visualisations, allowing clinicians to understand and trust the decision-making process. Comparative evaluation demonstrates that DT-ATTN outperforms both conventional ML and DL approaches, including SVM, RF, CNN, RNN, LSTM, and Transformer-based models. With its combination of accuracy, transparency, and real-time capability, the DT-ATTN framework demonstrates significant promise for deployment in computer-aided diagnosis, wearable cardiac monitoring systems, and telemedicine applications, ultimately contributing to enhanced patient care and informed clinical decision-making.

Footnotes

ORCID iDs

Imteyaz Hussain Khan

Amar Singh

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Alarsan

F. I.

Younes

(2019). Analysis and classification of heart diseases using heartbeat features and machine learning algorithms. Journal of Big Data, 6(81), 15. https://doi.org/10.1186/s40537-019-0244-x

Daydulo

Y. D.

Thamineni

B. L.

Dawud

A. A.

(2023). Cardiac arrhythmia detection using deep learning approach and time frequency representation of ECG signals. BMC Medical Informatics and Decision Making, 23(1), 1–14. https://doi.org/10.1186/s12911-023-02326-w

Denysyuk

H. V.

Pinto

R. J.

Silva

P. M.

Duarte

R. P.

Marinho

F. A.

Pimenta

Gouveia

A. J.

Gonçalves

N. J.

Coelho

P. J.

Zdravevski

Lameski

Leithardt

Garcia

N. M.

Pires

I. M.

(2023). Algorithms for automated diagnosis of cardiovascular diseases based on ECG data: A comprehensive systematic review. Heliyon, 9(2). https://doi.org/10.1016/j.heliyon.2023.e13601

Elsheikhy

D. H.

Hassan

A. S.

Yhiea

N. M.

Fareed

A. M.

Rashed

E. A.

(2025). Lightweight deep learning architecture for multi-lead ECG arrhythmia detection. Sensors, 25(7), 5542. https://doi.org/10.3390/s25175542

Feeny

A. K.

Chung

M. K.

Madabhushi

Attia

Z. I.

Cikes

Firouznia

Friedman

P. A.

Kalscheur

M. M.

Kapa

Narayan

S. M.

Noseworthy

P. A.

Passman

R. S.

Perez

M. V.

Peters

N. S.

Piccini

J. P.

Tarakji

K. G.

Thomas

S. A.

Trayanova

N. A.

Turakhia

M. P.

Wang

P. J.

(2020). Artificial intelligence and machine learning in arrhythmias and cardiac electrophysiology. Circulation: Arrhythmia and Electrophysiology, 13(8), E007952. https://doi.org/10.1161/CIRCEP.119.007952

Figuera

Irusta

Morgado

Aramendi

Ayala

Wik

Kramer-Johansen

Eftestøl

Alonso-Atienza

(2016). Machine learning techniques for the detection of shockable rhythms in automated external defibrillators. PLoS One, 11(7), e0159654-17. https://doi.org/10.1371/journal.pone.0159654

Ganesh Kumar

Kumaraswamy

Y. S.

(2014). Spline activated neural network for classifying cardiac arrhythmia. Journal of Computer Science, 10(9), 1582–1590. https://doi.org/10.3844/jcssp.2014.1582.1590

Hambarde

Paithane

Lambhate

Hambarde

A. S.

Kalyankar

P. A.

(2025). Smart arrhythmia detection using single lead ECG signal and hybridized deep neural network model. Web Intelligence, 23(2), 155–171. https://doi.org/10.1177/24056456241297300

Han

Yuan

Zhang

Liu

Panneerselvam

(2024). Novel transformation deep learning model for electrocardiogram classification and arrhythmia detection using edge computing. Journal of Grid Computing, 22(1), 1–19. https://doi.org/10.1007/s10723-023-09717-3

10.

Ismail

Serhani

M. A.

Hussein

N. M.

Elhadef

(2023). RL-ECGNet: Resource-aware multi-class detection of arrhythmia through reinforcement learning. . Applied Intelligence, 53(24), 30927–30939. https://doi.org/10.1007/s10489-023-05147-6

11.

Issa

M. F.

Yousry

Tuboly

Juhasz

Abuel-atta

A. H.

Selim

M. M.

(2023). Heartbeat classification based on single lead-II ECG using deep learning. Heliyon, 9(7), e17974. https://doi.org/10.1016/j.heliyon.2023.e17974

12.

Jat

Patil

Bhat

(2024). Detection of heart arrhythmia with electrocardiography. Network Modeling Analysis in Health Informatics and Bioinformatics, 13(1). https://doi.org/10.1007/s13721-024-00487-w

13.

Karthiga

Santhi

(2022). A wireless sensor network for remote detection of arrhythmias using convolutional neural network. Wireless Networks, 28(4), 1349–1360. https://doi.org/10.1007/s11276-021-02825-6

14.

Katal

Gupta

Verma

Sharma

(2023). Deep-learning-based arrhythmia detection using ECG signals: A comparative study and performance evaluation. Diagnostics, 13(24). https://doi.org/10.3390/diagnostics13243605

15.

Kim

Lee

K. R.

Lim

D. S.

Lee

K. H.

Lee

J. S.

Kim

D.-Y.

Sohn

C.-B.

(2025). A novel hybrid CNN-transformer model for arrhythmia detection without R-peak identification using stockwell transform. Scientific Reports, 15(1), 1–11. https://doi.org/10.1038/s41598-025-92582-9

16.

Lamba

Kumar

Diwakar

(2025). FADLEC: feature extraction and arrhythmia classification using deep learning from electrocardiograph signals. Discover Artificial Intelligence, 5(1). https://doi.org/10.1007/s44163-025-00290-0

17.

Madan

Singh

D. P.

Diwakar

Pant

Kishor

(2022). A hybrid deep learning approach for ECG-based arrhythmia classification. Bioengineering, 9(4), 152. https://doi.org/10.3390/bioengineering9040152

18.

Mani

Paul

T. S.

Archambault

P. M.

Marois

(2024). Machine learning workflow for edge computed arrhythmia detection in exploration class missions. npj Microgravity, 10(1). https://doi.org/10.1038/s41526-024-00409-0

19.

Mitra

Samanta

R. K.

(2013). Cardiac arrhythmia classification using neural networks with selected features. Procedia Technology, 10, 76–84. https://doi.org/10.1016/j.protcy.2013.12.339

20.

Mokesh Rayalu

Radhika

K. S. R.

Anitha Kumari

Botta

V. K. R.

Padmaja

Hussain

M. A.

(2024). ADAPT: An AI-driven approach for arrhythmia detection in diabetic patients. International Journal of System Assurance Engineering and Management, 0123456789, https://doi.org/10.1007/s13198-024-02619-x

21.

Mousavi

Afghah

(2019). Inter- and intra-patient ECG heartbeat classification for arrhythmia detection: A sequence to sequence deep learning approach. ICASSP, EEE International Conference on Acoustics, Speech, and Signal Processing - Proceedings, 2019 (pp. 1308–1312). https://doi.org/10.1109/ICASSP.2019.8683140

22.

Ramkumar

Alagarsamy

Balakumar

Pradeep

(2023). Ensemble classifier fostered detection of arrhythmia using ECG data. Medical & Biological Engineering & Computing, 61(9), 2453–2466. https://doi.org/10.1007/s11517-023-02839-6

23.

Reshad

A. I.

Nino

Valero

(2025). Deep learning-based detection of arrhythmia using ECG signals–A comprehensive review. Vascular Health and Risk Management, 21, 685–703. https://doi.org/10.2147/VHRM.S508620

24.

Ribeiro

Sá

Paiva

Rodrigues

P. M.

(2024). Cardiovascular diseases diagnosis using an ECG multi-band non-linear machine learning framework analysis. Bioengineering, 11(1), 58. https://doi.org/10.3390/bioengineering11010058

25.

Sai Kumar

Rinku

D. R.

Pradeep Kumar

Maddula

Anna Palagan

(2023). An IOT framework for detecting cardiac arrhythmias in real-time using deep learning resnet model. Measurement: Sensors, 29, 100866. https://doi.org/10.1016/j.measen.2023.100866

26.

Sheerinsithara

Raj

S. A. A.

(2025). ML-DLNN-based heartbeat detection system for diagnosing cardiac arrhythmia disease using ECG signal. Journal of Information Systems Engineering & Management, 10(9s), 248–264. https://doi.org/10.52783/jisem.v10i9s.1228

27.

Talukder

M. A.

Khalid

Kazi

Muna

N. J.

Nur-e-Alam

Halder

Sultana

(2024). A hybrid cardiovascular arrhythmia disease detection using ConvNeXt-X models on electrocardiogram signals. Scientific Reports, 14(1), 1–20. https://doi.org/10.1038/s41598-024-81992-w

28.

Tutuko

Passarella

Firdaus

Rachmatullah

M. N.

(2021). An automated system for ECG arrhythmia detection using machine learning techniques. Bioengineering (Basel), 10(1), 58.

29.

Ullah

S. I.

Ullah

Ishaq

Khan

Ghadi

Y. Y.

Algarni

(2024). Machine learning-based cardiovascular disease detection using optimal feature selection. IEEE Access, 12(0), 16431–16446. https://doi.org/10.1109/ACCESS.2024.3359910

30.

Zheng

Zhang

Danioko

Yao

Guo

Rakovski

(2020). A 12-lead electrocardiogram database for arrhythmia research covering more than 10,000 patients. Scientific Data, 7(1), 1–8. doi: 10.1038/s41597-020-0386-x