Comparative Evaluation of Deep Learning Models and Explainability Methods for Ecg Heartbeat Classification

Q: What is the clinical evidence from this study?

Population: Cardiac arrhythmias (ECG heartbeat classification) (n=109446). Intervention: Hybrid CNN-BiLSTM-Attention model vs. Compact 1D-CNN model. Primary outcome: Holdout test Macro F1 score for 5-class ECG heartbeat classification (MD -0.0541, 95% CI -0.0643 to -0.0440).

Q: What does this research mean for the field?

A compact 1D-CNN achieves a higher holdout Macro F1 score than a more complex hybrid CNN-BiLSTM-Attention model in multiclass ECG heartbeat classification. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

Jessica Fernandes Lopes; Felipe Torquato Bressan; Rafael S. Bressan; Danilo Sipoli Sanches

doi:10.47820/recima21.v7i10.9152

October 16, 2026RECIMA21 - Revista Científica Multidisciplinar - ISSN 2675-6218Open Access

Comparative Evaluation of Deep Learning Models and Explainability Methods for Ecg Heartbeat Classification

View Full Paper

Ask AI

Bookmark

Share

Key result

Compact 1D-CNN outperforms hybrid CNN-BiLSTM-Attention by ~5 points in multiclass ECG classification.

MD -0.0541
95% CI -0.0643 to -0.0440
n=109,446

Why the study?

This study was conducted to compare a 1D-CNN architecture and a hybrid CNN-BiLSTM-Attention model for multiclass ECG heartbeat classification while evaluating the faithfulness and temporal distribution of explainability techniques.

Does a hybrid CNN-BiLSTM-Attention model improve multiclass ECG heartbeat classification performance and explainability faithfulness compared to a compact 1D-CNN model in the MIT-BIH Arrhythmia Database?

Key Points

To compare a 1D-CNN architecture against a hybrid CNN-BiLSTM-Attention model for multiclass ECG heartbeat classification and evaluate the faithfulness and temporal distribution of explainability methods.
Evaluated models on an ECG dataset derived from the MIT-BIH Arrhythmia Database, categorized into Normal, Supraventricular, Ventricular, Fusion, and Unclassified classes.
Implemented five-fold stratified cross-validation on the training set with internal validation Macro F1 epoch selection, followed by final evaluation on an untouched holdout test partition.
Analyzed explainability using 1D Grad-CAM, Integrated Gradients, and temporal attention on a class-balanced subset of correctly classified heartbeats via a perturbation protocol.
On the holdout test set, the 1D-CNN attained a Macro F1 of 0.8681 compared with 0.8140 for the CNN-BiLSTM-Attention architecture, reversing the cross-validation ranking where the hybrid model led (0.8842 ± 0.0376 vs 0.8456 ± 0.0662).
The CNN-BiLSTM-Attention model achieved a higher test Balanced Accuracy than the 1D-CNN (0.9441 vs 0.9264), reflecting higher average class-wise recall.
Integrated Gradients yielded positive mean faithfulness gains across all classes for both architectures, whereas Grad-CAM and temporal attention showed class-dependent behavior with low-to-moderate agreement across methods.

Structured PICO

Does a hybrid CNN-BiLSTM-Attention model improve multiclass ECG heartbeat classification performance and explainability faithfulness compared to a compact 1D-CNN model in the MIT-BIH Arrhythmia Database?

P

Population

109,446 preprocessed 187-sample ECG heartbeats (87,554 training and 21,892 holdout test samples) from the MIT-BIH Arrhythmia Database across five arrhythmia categories.

I

Intervention

Hybrid CNN-BiLSTM-Attention deep learning model (150,149 trainable parameters, combining two convolutional layers with 64 and 128 filters, a Bidirectional LSTM layer with 64 hidden units per direction, and additive temporal attention) coupled with post-hoc Explainable AI (XAI) methods (Integrated Gradients and temporal attention).

C

Comparator

Compact 1D-CNN architecture (52,677 trainable parameters, composed of three convolutional blocks with 32, 64, and 128 filters, global average pooling, and a fully connected layer) coupled with 1D Grad-CAM and Integrated Gradients, evaluated under identical training partitions, weighted cross-entropy loss, and AdamW optimization.

O

Outcome

Macro F1 score on the untouched 21,892-sample holdout test set and cross-validation folds.surrogate

Main Result

Mean Difference: -0.0541 (95% CI -0.0643–-0.044)

Absolute Event Rate: 0.814% vs 0.8681%

Increased architectural complexity in a hybrid CNN-BiLSTM-Attention model did not yield uniformly superior predictive performance over a simpler 1D-CNN for multiclass ECG classification, while explainability analyses demonstrated that Integrated Gradients provided the most consistent behavioral faithfulness across classes.

Limitations

Preprocessed dataset files do not retain patient identifiers, preventing verification of patient-level independence between training and test partitions.
Absence of fiducial annotations in the preprocessed signals prevented direct mapping of attributions to physiological ECG landmarks such as P waves or QRS complexes.
Explainability analysis was restricted to a class-balanced subset of 500 heartbeats correctly classified by both architectures, which does not represent the entire holdout distribution.
Hyperparameters were fixed a priori for controlled comparison rather than tuned specifically for each architecture.
Preprocessed dataset files do not retain patient identifiers, so patient-level independence between training and test partitions cannot be established and within-patient dependence is not accounted for.
Hyperparameters (learning rate, weight decay, batch size) were fixed a priori and not globally optimized for either architecture.
Explainability analysis was restricted to a class-balanced subset of up to 500 holdout heartbeats correctly classified by both architectures, which does not represent the entire holdout distribution.
The all-zero reference baseline used for Integrated Gradients does not represent a physiologically neutral ECG.
Individual fiducial annotations (such as R peaks or QRS boundaries) are absent from the preprocessed representation and were not inferred.

Implication

Simpler 1D-CNNs may suffice over hybrids for multiclass ECG classification; leaves open optimal architecture selection and clinical translation.

Abstract

This study compared a 1D-CNN architecture and a hybrid CNN-BiLSTM-Attention model for multiclass electrocardiographic heartbeat classification, incorporating Explainable Artificial Intelligence techniques to evaluate the faithfulness and temporal distribution of the evidence used by the models. Experiments used a dataset derived from the MIT-BIH Arrhythmia Database, organized into the Normal, Supraventricular, Ventricular, Fusion, and Unclassified classes. The protocol employed five-fold stratified cross-validation exclusively on the training set, epoch selection based on Macro F1 in an internal validation partition, and final evaluation on an untouched holdout test set. The 1D-CNN achieved a test Macro F1 of 0.8681 versus 0.8140 for the CNN-BiLSTM-Attention model, whereas the hybrid model achieved higher Balanced Accuracy, 0.9441 versus 0.9264, indicating higher average class-wise recall. Explainability analysis employed 1D Grad-CAM, Integrated Gradients, and temporal attention on a class-balanced subset of samples correctly classified by both architectures. Under the adopted perturbation protocol, Integrated Gradients showed positive mean faithfulness gains across all classes for both architectures, whereas Grad-CAM and temporal attention exhibited more class-dependent behavior. Low-to-moderate agreement among the methods indicated that different techniques may emphasize distinct temporal regions for the same correct prediction. Within this experimental setting, the additional complexity of the hybrid model did not translate into uniformly better predictive performance, supporting the complementary evaluation of predictive performance and explainability in ECG classification.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Lopes et al. (2026) studied Cardiac arrhythmias (ECG heartbeat classification) (n=109,446). Hybrid CNN-BiLSTM-Attention model vs. Compact 1D-CNN model was evaluated on Holdout test Macro F1 score for 5-class ECG heartbeat classification (MD -0.0541, 95% CI -0.0643 to -0.0440). The compact 1D-CNN achieved a higher holdout Macro F1 score than the hybrid CNN-BiLSTM-Attention model (0.8681 vs 0.8140; difference -0.0541, 95% CI -0.0643 to -0.0440) in multiclass ECG classification.

synapsesocial.com/papers/6ae28e75cbadf70daf6c424c https://doi.org/10.47820/recima21.v7i10.9152

View Full Paper

Ask AI

Bookmark

Share

Comparative Evaluation of Deep Learning Models and Explainability Methods for Ecg Heartbeat Classification

Key Points

Structured PICO

Main Result

Limitations

Implication

Abstract

Discussion

Cite This Study

Comparative Evaluation of Deep Learning Models and Explainability Methods for Ecg Heartbeat Classification

Key Points

Structured PICO

Main Result

Limitations

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading