LightX3ECG: A Lightweight and eXplainable Deep Learning System for 3-lead Electrocardiogram Classification

This paper presents LightX3ECG, a novel lightweight and explainable deep learning system designed to accurately classify multiple cardiovascular abnormalities using only three ECG leads, thereby facilitating the use of portable and wearable devices for early detection.

The Gist

Imagine your heart is like a complex orchestra. To understand if the music is healthy or if a musician is playing out of tune, doctors usually listen to the entire orchestra using 12 different microphones (leads) placed all over the patient's body. This is the standard "12-lead ECG." It's the gold standard, but it's bulky, expensive, and requires a hospital visit.

Now, imagine you could get a just-as-good diagnosis using only three microphones attached to a simple, wearable wristband. That's exactly what the researchers behind LightX3ECG have built.

Here is the story of their invention, broken down into simple concepts:

1. The Problem: Too Much Gear, Not Enough Doctors

Heart disease is a massive global killer. Catching it early is key. But currently, analyzing heart signals (ECGs) is a slow, manual job that requires highly trained cardiologists. There aren't enough of them, especially in poorer regions. Plus, the machines that record the full 12-lead ECG are heavy and hard to carry around.

The researchers asked: Can we build a smart system that uses a tiny, portable device (with just 3 leads) to diagnose heart problems as accurately as the big hospital machines?

2. The Solution: The "Three-Headed Detective"

Most AI systems try to shove all 12 leads into one giant brain. But when you only have 3 leads, that approach is like trying to solve a puzzle with missing pieces.

Instead, the team built LightX3ECG, which works like a team of three specialized detectives:

• The Specialists: They use three separate "brains" (neural networks). Each brain looks at just one of the three heart signals (Lead I, Lead II, and Lead V1) on its own. This allows each brain to focus deeply on the unique patterns of that specific signal without getting confused by the others.
• The Team Leader (Attention Module): Once the three specialists have their opinions, they need to agree on a final verdict. The system uses a clever "Team Leader" module. This leader doesn't just average their opinions; it listens to them and decides, "Hey, Lead I is really important for this specific heart condition, so let's weigh its opinion more heavily." This creates a much smarter, more robust final decision.

3. The "Black Box" Problem: Why Trust the AI?

Deep learning AI is often called a "black box" because it gives an answer without explaining how it got there. In medicine, a doctor can't just say, "The computer said you have a heart attack," without knowing why. They need to see the evidence.

The "Flashlight" Technique:
To fix this, the researchers added a feature called Lead-wise Grad-CAM. Think of this as a flashlight that the AI turns on over the heart signal.

• When the AI diagnoses a problem, the flashlight highlights exactly where on the heartbeat graph the AI was looking.
• Did it see a weird spike in the first lead? Did it notice a missing wave in the second?
• The system shows three separate flashlights (one for each lead), giving the doctor a clear, visual map of why the diagnosis was made. It's like the AI saying, "I found the problem here, and here, and here."

4. The "Backpack" Problem: Making it Portable

AI models are usually huge, like a heavy backpack full of bricks. You can't fit that into a tiny wearable device or a phone.

The Pruning Trick:
The researchers used a technique called Pruning. Imagine you have a backpack full of tools. You realize that 80% of the tools are duplicates or rarely used. You throw them away. The backpack becomes 3 times lighter and easier to carry, but you can still do all the same jobs with the remaining essential tools.

• They cut out 80% of the "useless" math inside their system.
• The result is a tiny, lightweight system (only 6.5 MB!) that runs fast on simple hardware but still diagnoses heart issues with incredible accuracy.

5. The Results: Small Size, Big Smarts

They tested this system on two massive databases containing thousands of real patient records.

• Accuracy: It got an F1 score (a measure of accuracy) of 0.97 on one dataset and 0.80 on another. This is better than many existing, much larger systems.
• Efficiency: It uses far less computer power and storage space than its competitors.
• Trust: Doctors looked at the "flashlight" explanations and confirmed that the AI was looking at the right parts of the heart signal, just like a human expert would.

The Bottom Line

LightX3ECG is like a pocket-sized, super-smart heart doctor. It uses a clever team of three specialized AI brains, a smart leader to combine their thoughts, a flashlight to show its work, and a "lightweight" design that fits in your pocket. It proves that you don't need a massive hospital machine to get a world-class heart diagnosis; sometimes, three simple leads and a smart algorithm are all you need.

Technical Summary

1. Problem Statement

Cardiovascular diseases (CVDs) are a leading cause of global mortality. While the standard 12-lead Electrocardiogram (ECG) is the gold standard for diagnosis, it requires bulky clinical equipment and is often impractical for portable or wearable point-of-care devices. These devices typically capture only a subset of leads (often 1 to 3).

• The Challenge: Existing deep learning models for ECG classification are often designed for 12-lead inputs, are computationally heavy (large model size), and operate as "black boxes," lacking clinical interpretability.
• The Goal: Develop a system that accurately classifies cardiovascular abnormalities using only three leads (I, II, and V1), while being lightweight (suitable for edge devices) and explainable (providing clinical justification for predictions).

2. Methodology

The proposed system, LightX3ECG, utilizes a multi-input deep learning architecture combined with model compression and explainable AI (XAI) techniques.

A. Architecture Design

Instead of stacking the three leads into a single input tensor (common in 12-lead approaches), LightX3ECG treats each lead separately:

1. Three Separate Backbones: The system employs three distinct 1D-Squeeze-and-Excitation Residual Networks (1D-SEResNets).
   • Optimization: The standard 1D-SEResNet18 was redesigned for efficiency:
     • Larger Kernel Sizes: Expanded to capture longer temporal patterns in ECG signals.
     • Depth-wise Separable Convolutions (DSConv): Replaced standard convolutions to reduce parameters by ~80%.
2. Lead-wise Attention Module:
   • Features extracted from the three backbones are concatenated.
   • A learnable attention mechanism (FC layers + BatchNorm + Dropout + Sigmoid) calculates an importance score ($\alpha_i$) for each lead.
   • The final representation is a weighted sum of the three feature vectors based on these scores, allowing the model to dynamically prioritize the most informative leads for a specific prediction.
3. Classification Head: A Fully-Connected (FC) layer performs the final classification.

B. Explainable AI (XAI)

To address the "black box" issue, the authors propose Lead-wise Grad-CAM:

• Mechanism: Standard Grad-CAM generates Class Activation Maps (CAMs) from the gradients of the final convolutional layers.
• Innovation: The system generates three distinct CAMs (one for each lead). These maps are then multiplied by the attention scores ($\alpha_i$) derived from the Lead-wise Attention module.
• Result: This produces three weighted heatmaps overlaid on the input ECG signals, highlighting exactly which parts of which leads contributed most to the diagnosis, offering high clinical interpretability.

C. Model Compression (Pruning)

To ensure deployability on hardware-constrained platforms:

• Weight Pruning: A post-training technique was applied where 80% of the weights with the lowest $L_1$-norm (absolute value) were zeroed out.
• Outcome: This reduced the model size by approximately 3x (from ~19MB to ~6.5MB) with negligible loss in classification performance.

3. Key Contributions

1. Novel Architecture: Introduction of a multi-backbone 3-lead ECG classifier using 1D-SEResNets and a Lead-wise Attention mechanism, outperforming single-input stacking methods.
2. Explainability Framework: Development of Lead-wise Grad-CAM, which provides distinct, weighted activation maps for each input lead, bridging the gap between AI predictions and clinical reasoning.
3. Lightweight Deployment: Successful application of weight pruning to create a system small enough for wearable/portable devices without significant accuracy degradation.
4. State-of-the-Art Performance: Demonstrated superior performance on large-scale datasets compared to existing methods (e.g., 1D-ResNet34, TI-ResNet18).

4. Experimental Results

The system was evaluated on two large-scale public datasets: Chapman (10,646 recordings, multi-class) and CPSC-2018 (6,877 recordings, multi-label).

• Performance Metrics (F1 Scores):
  • Chapman Dataset: 0.9718 (Average).
  • CPSC-2018 Dataset: 0.8004 (Average).
  • Comparison: LightX3ECG outperformed state-of-the-art baselines (1D-ResNet34, 1D-SEResNet34, InceptionTime, TI-ResNet18) in both accuracy and efficiency.
• Efficiency:
  • Parameters: 5.31 Million (lowest among compared models).
  • FLOPs: 1.34 Billion.
  • Storage Size: 6.52 MB (after pruning), significantly smaller than competitors (e.g., 58.18 MB for 1D-ResNet34).
• Interpretability Validation:
  • Visual Check: The heatmaps correctly highlighted clinically relevant features (e.g., P-waves for Normal Sinus Rhythm, lack of P-waves for Atrial Fibrillation, ST-segment changes).
  • Sanity Check: A parameter randomization test confirmed that the explanations were dependent on the model's learned parameters (Spearman correlation of 0.10–0.11), proving the XAI technique is not random.

5. Significance

• Clinical Relevance: By focusing on leads I, II, and V1, the system aligns with the capabilities of modern wearable ECG devices, making advanced cardiac screening accessible in low-resource settings.
• Trust in AI: The integration of Lead-wise Grad-CAM addresses the critical barrier of trust in medical AI, allowing cardiologists to verify that the model is looking at the correct physiological markers.
• Edge AI Feasibility: The combination of architectural redesign (DSConv) and pruning demonstrates that high-performance medical AI can run on resource-constrained hardware, enabling real-time, on-device diagnosis.

In conclusion, LightX3ECG represents a significant step forward in making ECG analysis accurate, interpretable, and deployable outside of traditional hospital settings.