Wavelet-Domain Multi-Representation and Ensemble Learning for Automated ECG Analysis

This study demonstrates that an ensemble learning approach combining continuous wavelet transform-based time-frequency representations (scalograms and phasograms) with raw time-series data, optimized via weighted focal loss, achieves superior ECG classification performance (AUC 0.9233) on the PTB-XL dataset compared to individual modalities.

The Gist

Imagine your heart is a complex orchestra. When it's healthy, the musicians (your heart's electrical signals) play in perfect harmony. When something is wrong, the music gets a little off-key, or a specific instrument starts playing too loud or too soft.

Doctors use a machine called an ECG (Electrocardiogram) to record this "heart music" as a squiggly line on a graph. The problem is, reading these squiggly lines is like trying to understand a symphony just by looking at a single sheet of music; it's hard, it takes a long time, and different doctors might hear different things.

This paper is about teaching a computer to be a super-listener that can spot heart problems instantly and accurately. Here is how they did it, explained simply:

1. The Problem: The "Flat" View

Traditionally, computers looked at the ECG signal as just a flat line (Time Domain). It's like listening to a song only by the volume over time. You hear when the music is loud, but you miss the pitch and the texture that tell you if a violin is out of tune or a drum is broken.

2. The Solution: The "Spectrogram" Glasses

The researchers decided to give the computer special glasses called Wavelets. Instead of just looking at the line, these glasses turn the signal into a colorful map (a Scalogram) and a phase map (a Phasogram).

• The Scalogram (The Energy Map): Imagine taking a song and turning it into a heat map. Bright colors show where the energy is high. This helps the computer see how strong the heart's electrical beats are at different frequencies.
• The Phasogram (The Timing Map): This is like looking at the rhythm and the "phase" of the waves. It tells the computer how the waves are moving relative to each other.

By using both, the computer isn't just listening to the volume; it's seeing the entire musical score in 3D.

3. The Strategy: The "Tasting Panel"

The researchers didn't just trust one computer model. They set up a "tasting panel" (an Ensemble) with different experts:

• The Raw Signal Expert: A model that looks at the original squiggly line.
• The Energy Expert: A model that looks at the Scalogram (heat map).
• The Timing Expert: A model that looks at the Phasogram (rhythm map).

They tried two ways to combine these experts:

• Early Fusion (The Group Hug): They mashed the heat map and rhythm map together before feeding them to the computer. It's like giving the computer a single, super-detailed picture that has all the info mixed together.
• Late Fusion (The Committee Vote): They let each expert look at the data separately, make their own diagnosis, and then voted on the final answer.

4. The Twist: The "Weighted Score"

One big problem with medical data is that some heart problems are rare (like a specific type of heart attack), while "Normal" heartbeats are very common. If you train a computer on this, it gets lazy and just guesses "Normal" every time because it's usually right.

To fix this, the researchers used a Weighted Focal Loss. Think of this as a strict teacher who says, "If you get the rare, difficult cases wrong, you lose 5 points. If you get the easy 'Normal' cases wrong, you only lose 1 point." This forced the computer to pay extra attention to the rare, dangerous heart conditions.

5. The Result: The Super-Doctor

When they put it all together:

• The computer that looked at the raw line was good (92.2% accurate).
• The computers that looked at the colorful maps were also good.
• But the winner was the Team: By combining the raw line expert with the best map experts, and using the "Weighted Score" rule, they created a system that is 92.4% accurate.

Why This Matters

This isn't just about getting a higher score on a test. It means:

1. Faster Diagnosis: The computer can analyze a heart signal in about 22 milliseconds (faster than a human blink).
2. Fewer Missed Cases: By focusing on the rare diseases, it helps catch heart problems that humans might miss because they are too busy or tired.
3. Better Tools: It proves that looking at heart signals as "images" (maps) rather than just "lines" is a powerful way to understand heart health.

In short, the researchers built a digital detective that uses special glasses to see the hidden patterns in your heart's music, ensuring that even the quietest, rarest warning signs don't get ignored.

Technical Summary

1. Problem Statement

Cardiovascular diseases (CVDs) are the leading cause of global mortality, and Electrocardiogram (ECG) analysis is the primary non-invasive diagnostic tool. However, manual interpretation is time-consuming, prone to inter-observer variability, and difficult to scale. While Deep Learning (DL) models have improved automated classification, most existing approaches rely solely on time-domain waveforms, overlooking critical time-frequency information that characterizes pathological patterns.

Furthermore, the PTB-XL dataset (a large, publicly available 12-lead ECG dataset with multi-class and multi-label annotations) has not been extensively explored using Continuous Wavelet Transform (CWT) representations (scalograms and phasograms) combined with advanced fusion strategies for multi-class, multi-label diagnostic tasks. The specific challenge addressed is the classification of ECGs into five diagnostic super-classes (Normal, Myocardial Infarction, ST/T Change, Conduction Disturbance, and Hypertrophy) while handling class imbalance and leveraging complementary spectral and phase data.

2. Methodology

The authors propose a comprehensive framework involving data preprocessing, multi-representation generation, and diverse modeling strategies.

A. Data Preprocessing & Representation

• Dataset: PTB-XL (21,837 12-lead recordings, 10s duration, 500/100 Hz).
• Signal Processing:
  1. Filtering: Band-pass filtering (0.05–47 Hz) to remove baseline wander and noise.
  2. Normalization: Z-score standardization per lead.
  3. Continuous Wavelet Transform (CWT): Applied to extract time-frequency features using a Morlet wavelet.
• Multi-Representation Generation:
  • Scalogram: The magnitude (amplitude) map representing spectral energy.
  • Phasogram: The phase map representing phase distribution.
  • Format: Both are normalized to [0,1], resized to 224×224 pixels, and treated as 12-channel images (one channel per ECG lead).

B. Modeling Strategies

The study evaluates several architectures and fusion techniques:

1. Time-Domain Baseline:
   • Model: XResNet1d101 (1D CNN) processing raw ECG signals.
2. Time-Frequency Domain (Single Modality):
   • Custom CNN: CWT2DCNN designed to natively handle 12-channel inputs.
   • Transfer Learning: Fine-tuning pretrained image models (ResNet50, EfficientNetB0, Swin Transformer). Since these expect 3 channels, a $1\times1$ convolutional layer maps the 12 leads to 3 channels.
   • Hybrid Models: CNN-Swin Transformer architectures.
3. Fusion Strategies:
   • Early Fusion: Concatenating scalogram and phasogram tensors into a 24-channel input before feeding into the network (allowing joint spectral-phase learning).
   • Late Fusion: Training independent models for scalograms and phasograms, then concatenating their embeddings (512-dim each) into a 1024-dim vector for final classification via an MLP.
4. Loss Functions:
   • Standard Binary Cross-Entropy (BCE).
   • Weighted Focal Loss: Implemented to address class imbalance and improve detection of minority labels.

3. Key Contributions

• Novel Multi-Representation Framework: First study to systematically apply and compare scalograms and phasograms derived from the PTB-XL dataset for multi-class, multi-label ECG classification.
• Fusion Strategy Comparison: A rigorous evaluation of Early vs. Late fusion strategies, demonstrating that early fusion (joint learning of magnitude and phase) generally yields superior feature extraction.
• Hybrid Architecture Integration: Successful adaptation of Swin Transformers and Hybrid CNN-Transformer models for 12-lead ECG wavelet images, showing they outperform standard CNNs in this domain.
• Ensemble Optimization: Development of an ensemble model combining the best time-domain baseline (XResNet1d101) with the best early-fusion wavelet models, achieving state-of-the-art performance on the dataset.
• Imbalance Handling: Demonstration that Weighted Focal Loss significantly improves classification rates for underrepresented diagnostic classes.

4. Experimental Results

Experiments were conducted on the PTB-XL test split (2,163 samples) using macro AUC, F1, and F$\beta$ metrics.

• Baseline Performance: The raw time-domain model (XResNet1d101) achieved a strong baseline AUC of 0.9224.
• Single Modality (Wavelet):
  • Hybrid Swin Transformers performed best on both scalograms (AUC 0.8819) and phasograms (AUC 0.8712).
  • Single modalities generally underperformed the raw time-domain baseline.
• Fusion Performance:
  • Early Fusion outperformed Late Fusion across all architectures. The Hybrid Swin Transformer with Early Fusion achieved an AUC of 0.9018.
• Ensemble & Weighted Loss (Best Result):
  • The final ensemble, combining the XResNet1d101 (time-domain) with the best Early Fusion models (Hybrid Swin and ResNet50) trained with Weighted Focal Loss, achieved the highest overall Macro AUC of 0.9238.
  • This surpassed the individual time-domain baseline and all single-modality wavelet models.
• Efficiency:
  • Inference time for the raw model: 1.412 ms/sample.
  • Inference time for the ensemble: 22.335 ms/sample.
  • Despite higher latency, the ensemble remains feasible for near real-time clinical deployment.

5. Significance and Conclusion

This research validates that time-frequency representations (specifically CWT scalograms and phasograms) provide complementary information to raw time-domain signals, enhancing the robustness of automated ECG diagnosis.

• Clinical Impact: The proposed framework offers a more interpretable and generalizable approach to detecting complex cardiac abnormalities, potentially reducing diagnostic errors and variability.
• Technical Insight: The study highlights that Early Fusion of magnitude and phase data allows models to learn joint spectral-phase representations more effectively than Late Fusion.
• Future Directions: The authors suggest exploring multi-head classification, generative augmentation to further mitigate data imbalance, and lead-subset analysis to optimize model efficiency for resource-constrained environments.

In summary, the paper demonstrates that a multi-representation wavelet fusion strategy combined with ensemble learning and focal loss represents a significant advancement in automated ECG analysis, setting a new benchmark for the PTB-XL super-diagnostic task.