⚡ electrical engineering

Deep Neural Network Architectures for Electrocardiogram Classification: A Comprehensive Evaluation

This study evaluates and compares various deep neural network architectures for ECG arrhythmia classification, demonstrating that a dynamic Top2-Weighted ensemble of CNN-LSTM and ResNet-1D models, augmented by GAN-generated data, achieves the highest performance (F1-score of 0.958) while ensuring interpretability through Grad-CAM.

Original authors: Yun Song, Wenjia Zheng, Tiedan Chen, Ziyu Wang, Jiazhao Shi, Yisong Chen

Published 2026-02-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yun Song, Wenjia Zheng, Tiedan Chen, Ziyu Wang, Jiazhao Shi, Yisong Chen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your heart is a drummer in a band. Sometimes, the drummer keeps a perfect, steady beat. Other times, they might skip a beat, rush ahead, or get confused by a different rhythm. Doctors use a tool called an ECG (electrocardiogram) to listen to this "drummer" and diagnose if something is wrong.

However, listening to thousands of hours of heartbeats manually is like trying to find a needle in a haystack while wearing blindfolds. It's tiring, slow, and humans make mistakes.

This paper is about building a super-smart AI assistant to listen to these heartbeats for us. The researchers didn't just build one assistant; they built four different types of "detectives" and tested them to see which one is best at spotting heart rhythm problems (arrhythmias).

Here is the story of their experiment, broken down simply:

1. The Problem: The "Unbalanced Class"

Imagine a classroom where 90% of the students are wearing blue shirts (Normal heartbeats), and only a few are wearing red, green, or yellow shirts (Rare heart problems).
If you ask a teacher to spot the red-shirted students, they might get lazy and just guess "Blue" for everyone because they are so common. The AI faces the same problem: it sees too many "Normal" beats and not enough "Sick" beats.

The Fix: The researchers used a magical photocopier called a GAN (Generative Adversarial Network). This AI didn't just copy the rare beats; it invented new, realistic fake beats that looked exactly like the rare ones. This balanced the classroom, forcing the AI to pay attention to the minority groups.

2. The Four Detectives (The Models)

The team trained four different AI "detectives," each with a unique way of looking at the heartbeat:

Detective CNN (The Microscope): This detective looks at the heartbeat very closely, focusing on the shape of individual waves (like the QRS complex). It's great at seeing what the beat looks like right now, but it doesn't remember what happened a second ago.
Detective CNN-LSTM (The Storyteller): This detective uses a microscope and has a great memory. It looks at the shape of the beat but also remembers the rhythm of the previous beats. It understands the "story" of the heart's rhythm over time.
Detective CNN-LSTM-Attention (The Highlighter): This is the Storyteller with a highlighter pen. It tries to focus only on the most important parts of the heartbeat and ignore the rest. The idea was good, but in practice, it got a bit confused when the rare beats were tricky.
Detective ResNet (The Deep Diver): This detective goes very deep into the layers of the signal, trying to find hidden patterns. However, it was a bit unstable, sometimes getting lost in the deep end.

3. The Results: Who Won?

When they tested them on a "final exam" (a test set of heartbeats):

The Microscope (CNN) was good at spotting local weirdness but missed the bigger picture.
The Highlighter (Attention) and The Deep Diver (ResNet) got confused by the rare heartbeats. They were great at spotting common problems but failed miserably when the rare, tricky ones showed up.
The Storyteller (CNN-LSTM) was the champion of the single detectives. It struck the perfect balance. It could see the shape of the beat and understand the rhythm over time. It was the most reliable on its own.

4. The Super-Team: Ensemble Learning

The researchers realized that even the best detective makes mistakes. So, they decided to form a Super-Team.

They took the top two detectives (The Microscope and The Storyteller) and asked them to vote together.

Equal Vote: They just averaged their answers.
Weighted Vote: They let the better detective's opinion count slightly more, but still listened to the other one.

The Winner: The Weighted Super-Team (specifically the "Top2-Weighted" ensemble) achieved the highest score of 0.958.

Think of it like a medical board meeting. If one doctor says "It's a heart attack" and another says "It's just a skipped beat," the AI listens to both, weighs their expertise, and makes a final, highly accurate decision. This team approach made the system much harder to fool.

5. Why This Matters

This paper proves that:

Context is King: To understand a heartbeat, you need to see the shape and the rhythm over time (hence why the Storyteller won).
Teamwork Makes the Dream Work: Combining different AI models is better than relying on just one "super" model.
Fairness: By using the magical photocopier (GAN) to balance the data, the AI learned to treat rare heart conditions just as seriously as common ones.

In a nutshell: The researchers built a smart, balanced team of AI detectives that can listen to your heart, understand its story, and spot dangerous rhythms with near-perfect accuracy. This brings us one step closer to AI that can help doctors save lives faster and more reliably.

1. Problem Statement

Cardiovascular diseases remain a leading cause of global mortality, necessitating early detection via Electrocardiogram (ECG) monitoring. While deep learning has revolutionized automated arrhythmia detection, existing research suffers from several critical gaps:

Inconsistent Evaluation: Studies often compare models under varying data processing and training conditions, making fair architectural comparisons difficult.
Class Imbalance: Real-world ECG datasets (like MIT-BIH) are heavily skewed toward normal heartbeats, causing models to underperform on minority arrhythmia classes (e.g., Atrial Premature beats).
Architectural Trade-offs: It remains unclear which specific architectural components (local feature extraction vs. temporal modeling vs. attention mechanisms) offer the best balance between sensitivity, specificity, and robustness under severe data imbalance.
Lack of Ensemble Optimization: Few studies systematically investigate how fusing complementary architectures can mitigate individual model weaknesses.

2. Methodology

The study proposes a unified evaluation framework to compare four distinct deep learning architectures on the MIT-BIH Arrhythmia Database.

A. Data Preprocessing and Augmentation

Dataset: Utilized the MIT-BIH dataset (48 half-hour recordings, 360 Hz sampling).
Classification Task: Mapped annotations into five clinically significant classes: Normal (N), Atrial Premature (A), Premature Ventricular Contraction (V), Fusion of Paced and Normal (f), and Fusion of Ventricular and Normal (F).
Preprocessing: Implemented heartbeat segmentation (188 samples centered on R-peaks), Min-Max normalization, and stratified splitting (85% train, 15% validation).
GAN-based Augmentation: To address severe class imbalance, a Sequence-Aware Generative Adversarial Network (GAN) was developed.
- Architecture: Both Generator and Discriminator use Bidirectional LSTMs to preserve temporal continuity.
- Process: The GAN synthesizes minority-class beats, which are filtered by discriminator confidence before merging with real data to create a balanced training set.

B. Model Architectures

Four distinct architectures were designed and compared:

CNN: A hierarchical 1D Convolutional Neural Network using ConvNormPool blocks (Conv + BatchNorm + Swish + Residual) to capture local morphological features (QRS, P-waves).
CNN-LSTM: Integrates the CNN feature extractor with a Bidirectional LSTM layer to model long-term temporal dependencies between consecutive heartbeats.
CNN-LSTM-Attention: Adds an attention mechanism atop the LSTM outputs to dynamically weight diagnostically salient temporal regions, aiming to improve interpretability.
ResNet-1D: A 1D Residual Network using deep residual blocks with identity shortcuts to facilitate gradient flow and hierarchical feature learning.

C. Training Strategy

Loss Function: Focal Loss ( $\alpha=1, \gamma=2$ ) was used to down-weight easy examples and focus training on hard, minority-class samples.
Optimizer: AdamW with decoupled weight decay for better generalization.
Scheduler: ReduceLROnPlateau to dynamically adjust learning rates based on validation loss stagnation.
Ensemble Learning: A late-fusion strategy was employed, aggregating outputs via weighted averaging. Strategies included Equal-Weight (All, Top3, Top2) and Top2-Weighted (weights proportional to validation macro-F1 scores).

D. Evaluation Metrics

Performance was assessed using Accuracy, Precision, Recall, F1-score, and AUC-ROC. 95% Confidence Intervals (CIs) were calculated to ensure statistical robustness. Grad-CAM was used for model interpretability to verify if models attended to physiologically relevant regions (e.g., QRS complexes).

3. Key Contributions

Unified Benchmark: Established a fair comparison framework for CNN, CNN-LSTM, CNN-LSTM-Attention, and ResNet-1D under identical preprocessing and training conditions.
Sequence-Aware GAN: Developed a GAN specifically for ECG data augmentation that preserves physiological waveform fidelity while balancing minority classes.
Ensemble Fusion Design: Introduced and evaluated performance-weighted ensemble strategies, demonstrating that fusing complementary models outperforms single architectures.
Empirical Insights: Provided a systematic analysis of how architectural choices (temporal modeling vs. attention) impact robustness against class imbalance.

4. Experimental Results

Individual Model Performance

CNN-LSTM: Achieved the best balance among single models with an F1-score of 0.951. It demonstrated the most stable convergence and consistent performance across all classes, effectively capturing both local morphology and temporal context.
CNN: Showed strong local feature extraction and high recall for Atrial Premature (A) beats (0.82) but lacked long-term temporal modeling.
CNN-LSTM-Attention & ResNet-1D: Exhibited high variability and instability.
- The Attention model collapsed on the minority A class (Recall $\approx$ 0.03, F1 $\approx$ 0.05), suggesting that attention mechanisms can amplify class imbalance issues without sufficient data.
- ResNet-1D showed the widest variance in confidence intervals and the lowest macro-average AUC (0.892).

Ensemble Performance

The Top2-Weighted Ensemble (combining CNN and CNN-LSTM based on their validation F1 scores) achieved the highest overall performance:
- F1-score: 0.958
- Precision: 0.986
- Recall: 0.934
This configuration successfully mitigated the weaknesses of individual models, improving sensitivity to minority arrhythmias while maintaining high precision.

Interpretability (Grad-CAM)

Visualizations confirmed that the CNN-LSTM model correctly focused on clinically relevant regions (steep depolarization peaks in QRS complexes and ST-T segments for abnormalities), validating its decision-making process.

5. Significance and Conclusion

This study demonstrates that temporal modeling is critical for robust ECG classification. While CNNs excel at local feature extraction, the integration of LSTM layers provides the necessary context to handle sequential dependencies, leading to superior generalization.

Key Finding: Complex additions like Attention mechanisms or deep Residual structures do not automatically guarantee better performance; in fact, they can introduce instability and sensitivity to class imbalance if not carefully tuned.
Practical Impact: The proposed Top2-Weighted Ensemble framework offers a reliable, interpretable, and high-performance solution for intelligent arrhythmia detection systems. It provides a robust foundation for clinical deployment, particularly in scenarios where data imbalance is a significant challenge.
Future Directions: The authors suggest exploring adaptive attention mechanisms, further synthetic data generation for rare classes, and the development of lightweight models for real-time deployment.

In summary, the paper establishes that a hybrid approach combining CNNs and LSTMs, augmented by GAN-based balancing and performance-weighted ensemble fusion, represents the state-of-the-art approach for automated, reliable ECG arrhythmia classification.