⚡ electrical engineering

Validating the Clinical Utility of CineECG 3D Reconstructions through Cross-Modal Feature Attribution

This paper proposes and validates a cross-modal method that projects feature attributions from high-performance 12-lead ECG deep learning models onto CineECG 3D anatomical space, demonstrating that this approach significantly improves the clinical interpretability and localization of pathological features compared to standard attribution techniques.

Original authors: Karol Dobiczek, Maciej Mozolewski, Szymon Bobek, Michał Szafarczyk, Peter van Dam, Grzegorz J. Nalepa

Published 2026-05-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Karol Dobiczek, Maciej Mozolewski, Szymon Bobek, Michał Szafarczyk, Peter van Dam, Grzegorz J. Nalepa

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

The Big Picture: The "Black Box" Problem

Imagine a super-smart robot doctor that can look at a standard 12-lead electrocardiogram (ECG)—the squiggly lines on a piece of paper—and tell you if a heart is sick with 99% accuracy. It's a genius at diagnosing.

But there's a catch: the robot is a "black box." It gives you the answer, but it can't explain why. It's like a student who gets every math problem right but can't show their work. Doctors can't trust a machine if they don't understand its reasoning.

Usually, when we try to make these robots explain themselves, we get a messy heatmap showing which parts of the squiggly lines mattered. But translating those abstract lines into actual heart anatomy (like "the left ventricle is acting up") is incredibly hard. It's like trying to describe the location of a fire in a house just by looking at the smoke coming out of the chimney, without seeing the house itself.

The Solution: A 3D Map (CineECG)

The researchers introduced a new tool called CineECG. Think of this as a 3D GPS map of the heart's electrical activity. Instead of just looking at flat lines, CineECG shows the electrical signal traveling through the heart's actual 3D shape, like a glowing path moving through a model of a heart.

The Experiment: Two Approaches

The team wanted to see if they could combine the "genius robot" (the high-accuracy model) with the "3D map" (CineECG) to make the explanation clear. They tested two main strategies:

The "Native" Approach: They trained a robot to learn only from the 3D map.
- The Result: This was a failure. The robot got less accurate at diagnosing, and its explanations were confusing and scattered. It was like trying to teach a student to drive using only a video game; they might understand the game, but they can't handle the real car.
The "Cross-Modal" Approach (The Winner): They trained the robot on the standard 12-lead lines (where it is a genius) and then projected its reasoning onto the 3D map.
- The Analogy: Imagine the robot is a detective solving a crime using a list of clues (the 12-lead lines). Instead of just showing the list to the police, the detective takes those clues and pins them onto a physical 3D model of the city where the crime happened. Now, the police can instantly see where the clues point in real space.

What They Found

The researchers tested this "Cross-Modal" method against a group of expert cardiologists who acted as the "ground truth" (the gold standard).

Better Accuracy in Explanation: When the robot explained its diagnosis by projecting its thoughts onto the 3D map, it matched the experts' thinking 56% of the time (measured by a "Dice score").
The Baseline: When they just looked at the standard 12-lead lines without the 3D map, the match was only 47%.
The "Magic" Filter: The paper claims that this projection acts like a noise filter. Sometimes, the robot's internal logic gets shaky or confusing (like static on a radio). By averaging the data and projecting it onto the smooth 3D path, the "static" disappears, and the true signal (the actual heart problem) becomes clear and easy to spot.

A Specific Example: The "Left Anterior Fascicular Block"

To prove it worked, they looked at a specific heart condition called LAFB.

The Native 3D Model: When the robot tried to explain this using only the 3D data, it got confused and focused on the very beginning of the heartbeat (the start of the signal), which wasn't actually the problem.
The Mapped Model: When they used the "Cross-Modal" method, the robot correctly ignored the start and pointed the 3D map exactly to the part of the heart where the electrical block was happening. It successfully translated the "what" (the diagnosis) into the "where" (the anatomy).

The Conclusion

The paper concludes that you don't need to throw away the powerful 12-lead models to get good explanations. Instead, you can keep the powerful model and simply translate its output onto the 3D heart map.

This creates a system that is:

Accurate: It keeps the high diagnostic power of the standard model.
Intuitive: It gives doctors a clear, 3D visual of where the heart is sick, rather than a confusing list of numbers.

In short, they found a way to make the robot doctor "show its work" in a language (3D anatomy) that human doctors can actually understand and trust.

1. Problem Statement

While deep learning models for 12-lead Electrocardiogram (ECG) analysis achieve high diagnostic accuracy, they suffer from a lack of interpretability, which hinders clinical adoption. Standard Explainable AI (XAI) methods (e.g., feature attribution heatmaps) struggle to map abstract temporal waveform fluctuations to specific physical anatomical pathologies.

The Gap: Clinicians require spatial reasoning (locating where in the heart a problem exists), but standard ECG models provide only temporal focus (identifying when in the signal a feature occurs).
The Limitation of Direct 3D Training: Previous attempts to train models directly on 3D anatomical signals (CineECG) result in reduced diagnostic accuracy and incoherent, unstable feature attributions.
The Challenge: How to combine the high predictive power of standard 12-lead models with the intuitive anatomical clarity of CineECG 3D reconstructions without sacrificing performance.

2. Methodology

The authors propose a Cross-Modal Feature Attribution Mapping framework. Instead of training a model directly on 3D data, they use a high-performance 1D model trained on standard 12-lead ECGs and project its explanations onto the 3D CineECG space.

A. Data and Architecture

Dataset: A synchronized subset of the PTB-XL dataset (8,000 records, balanced Normal/Abnormal) with strict temporal alignment between 12-lead ECG and CineECG signals. An independent cohort of 20 expert-annotated cases was used for ground-truth validation.
Model: A custom 1D-ResNet architecture designed for strict differentiability (essential for gradient-based XAI). It avoids pooling operations that cause "shattered gradients," using strided convolutions and replication padding instead.
Training: The model was trained to classify Normal vs. Abnormal signals. It achieved 83.06% accuracy on 12-lead data and 79.06% on direct CineECG data.

B. Cross-Modal Mapping Pipeline

Feature Extraction: The 12-lead ECG model generates raw feature attributions ( $A$ ) for each lead and time-step.
Bipolar Normalization: Raw attributions are converted into a unified bipolar importance profile ( $\Phi$ ) to distinguish evidence for abnormality (positive) vs. normality (negative).
Aggregation: The 12-lead attributions are aggregated into a global temporal vector.
Projection: This aggregated temporal vector is uniformly applied to all spatial dimensions ( $x, y, z$ ) of the CineECG ventricular model at each time-step, effectively "coloring" the 3D trajectory based on the 12-lead model's focus.

C. Evaluation Protocol

Ground Truth: Two domain experts independently annotated the 20-case validation set, marking salient pathological segments on both 12-lead ECGs and 3D CineECG trajectories.
Metrics: Agreement was measured using:
- Dice Coefficient & IoU: To measure spatial overlap between model attributions and expert masks.
- Spearman Correlation: To evaluate the rank-order alignment of attribution values with expert judgments.
XAI Methods Tested: Gradient-based (Integrated Gradients, Gradient SHAP) and Perturbation-based (LIME, Kernel SHAP).

3. Key Contributions

Novel Cross-Modal Framework: The first method to project feature attributions from a high-performance 1D ECG model onto a 3D CineECG anatomical space, bridging the gap between algorithmic complexity and clinical spatial reasoning.
Validation of 3D Utility: Demonstrated that while training directly on CineECG degrades model performance and explanation coherence, using CineECG as a visualization layer for 12-lead models significantly improves clinical interpretability.
Optimization of XAI Strategies: Identified that Integrated Gradients (IG) combined with Scaled normalization is the optimal configuration for this cross-modal mapping, outperforming perturbation-based methods (LIME/SHAP) which failed to maintain meaningful rank correlations in high-dimensional spatial data.
Diagnostic Debugging Tool: Revealed that the mapping mechanism can successfully localize pathological features even when the classification head fails (i.e., when the model predicts the wrong label), making it a valuable tool for verifying network focus.

4. Results

Superior Alignment: The cross-modal mapping (12-lead attributions projected to CineECG) achieved a Dice score of 0.56, significantly outperforming the standard 12-lead baseline (0.47) and the direct CineECG model attribution (0.49).
Rank Correlation: The mapped approach achieved a Spearman correlation of 0.22, compared to 0.18 for the 12-lead baseline and a negligible 0.03 for direct CineECG attributions. This proves that direct 3D training fails to preserve the ranking of clinically relevant features.
Robustness to Errors: In cases where the model's diagnosis disagreed with the expert's diagnosis, the mapped attributions still achieved a high Dice score (0.57), indicating the model "saw" the pathology correctly even if it misclassified the case.
Qualitative Findings:
- Direct CineECG explanations often focused erroneously on signal onset (0–25 ms) rather than the pathology.
- Mapped explanations correctly localized ventricular depolarization (50–100 ms) for conditions like Left Anterior Fascicular Block (LAFB).
- The mapping acts as a spatial regularizer, filtering out "vanishing gradient" noise that occurs at isoelectric lines in 1D signals.

5. Significance and Conclusion

This study establishes that the true clinical value of CineECG lies not in replacing standard ECG models, but in contextualizing their outputs. By projecting the robust diagnostic focus of 12-lead models onto a 3D anatomical frame, the method:

Reduces Cognitive Load: Transforms twelve abstract temporal plots into a single, intuitive 3D visualization.
Enhances Trust: Provides clinicians with a spatially grounded explanation that aligns with anatomical knowledge.
Filters Instability: The averaging mechanism inherent in the mapping smooths out the noise and instability often found in gradient-based attributions.

The authors conclude that cross-modal averaging mapping is a critical step toward integrating deep learning into clinical cardiology, offering a solution that combines the diagnostic expressiveness of standard ECG with the intuitive clarity of anatomical visualization.