💻 computer science

Integrated Gradient-guided Adversarial Counterfactual Explanations of 12-lead Electrocardiograms

This paper introduces InterGRACE, an Integrated Gradient-guided adversarial framework that generates clinically meaningful counterfactual explanations for 12-lead ECGs by targeting decision-relevant features like the ST segment and T wave, thereby enhancing the interpretability of AI models for myocardial infarction detection.

Original authors: Bjørn-Jostein Singstad

Published 2026-09-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Bjørn-Jostein Singstad

Original paper licensed under CC BY 4.0 (https://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

Every day, millions of people walk into hospitals with chest pain or shortness of breath, and the first tool doctors reach for is often a simple, ten-second recording of the heart's electrical activity. This is the electrocardiogram, or ECG, a test that maps the heart's rhythm across twelve different angles, revealing hidden dangers like heart attacks long before symptoms become severe. For decades, this test has been the gold standard, but today, artificial intelligence is learning to read these same squiggly lines with a speed and accuracy that sometimes rivals human experts. Yet, a quiet barrier remains between the computer and the clinic: while a machine can tell a doctor that a patient is at risk, it often cannot explain why. It sees a pattern in the data that a human cannot see, leaving the physician to trust a black box. To bridge this gap, researchers are developing a new way of thinking about how machines make decisions, moving beyond simply pointing out which part of the image looked important, to asking a more human question: what would have to change for the answer to be different?

This line of thinking, known as counterfactual explanation, mirrors the way doctors actually diagnose patients. When a physician looks at a heart tracing, they do not just identify features; they mentally simulate alternatives. They ask themselves, if this wave were slightly higher, or if this pause were shorter, would the diagnosis shift from a heart attack to a normal rhythm? This kind of reasoning is central to medical practice, but it has been difficult to teach to computers. Most existing methods for making artificial intelligence transparent can highlight the regions of an ECG that influenced a prediction, much like a heat map showing where a camera focused. However, these maps do not show the specific, minimal change required to flip the result. They tell you where the machine looked, but not what it would take to change its mind.

In a new study, a researcher named Bjørn-Jostein Singstad set out to build a system that could generate these "what-if" scenarios for heart signals. The goal was to create a framework that could take an ECG classified as a heart attack and gently nudge it, in a way that concentrates on the specific parts of the signal the computer itself deemed most important, until the computer saw it as a normal heart, and vice versa. The challenge was to ensure that these changes were not random noise or digital artifacts, but modifications that focused on the critical moments the machine relied on. To do this, the researcher developed a method called InterGRACE. This system uses a technique called integrated gradients, which acts like a spotlight, identifying exactly which moments in the heartbeat the machine relies on most for its decision. The system then uses this spotlight to guide a process that alters the signal, focusing the changes on those critical moments while keeping the rest of the wave as smooth and natural as possible.

The study tested this approach on a large, public collection of over twenty-one thousand heart recordings, focusing specifically on distinguishing between heart attacks and normal rhythms. The researcher trained two separate computer models to read these signals, ensuring they learned the task independently of one another. The goal was to see if the changes suggested by the system were genuine medical features or just quirks of a single computer program. When the system was asked to flip the diagnosis of a heart attack to a normal rhythm, it succeeded in changing the mind of the first computer model in nearly two-thirds of the cases. More importantly, when those same altered signals were shown to the second, independently trained computer, it also changed its mind in more than half of the cases. This partial success suggests that the system is finding some features that are truly shared between different ways of thinking about the data, while also indicating that a large fraction of the changes remain specific to the originating model.

Looking closely at the signals that were changed, a clear pattern emerged. The system did not distort the entire heartbeat randomly. Instead, it concentrated its modifications on a specific window of time known as the ST segment and the T wave. In medical terms, these are the parts of the heartbeat that occur just after the main electrical spike, representing the heart's recovery phase. This is significant because, in real-world medicine, changes in these exact areas are the primary indicators of a heart attack. The fact that the computer, guided only by its own internal logic, chose to alter these specific regions suggests it aligns with the presence of clinically meaningful repolarization-related features, although the study notes this does not conclusively demonstrate that these are the only features at play. The changes were not uniform across the whole signal; they were precise, targeting the recovery phase while leaving the initial electrical spike largely untouched.

However, the study also revealed important limits to this technology. While the system worked well for the computer models, the changes it produced were not always perfectly safe or realistic for a human to interpret. Because the system treated each of the twelve leads of the ECG independently, it sometimes created combinations of signals that could never occur in a living human heart. For instance, it might change the voltage in one lead without making the corresponding, physically necessary change in another lead, creating a pattern that violates the laws of how electricity travels through the body. The researcher noted that without strict rules to enforce these biological laws, the generated examples can contain features that are unlikely or impossible in real cardiac signals, such as sudden jumps in voltage or distorted waves that no real heart could produce.

This finding serves as a crucial reminder that while the method successfully identifies which parts of the signal matter to the machine, it does not yet fully understand the complex physics of the human body. The study concludes that this approach is a promising step toward making artificial intelligence more transparent, offering a way to probe the reasoning of a computer model by asking what it would take to change its mind. Yet, before such tools can be used to help doctors make life-or-death decisions, they must be refined to ensure that every suggested change is not only effective for the computer but also physically possible for a human heart. The work stands as an initial demonstration that machines can be guided to think in terms of "what if," but it also highlights the long road ahead to ensure those hypothetical changes are grounded in the reality of human physiology.

Technical Summary: Integrated Gradient-guided Adversarial Counterfactual Explanations of 12-lead Electrocardiograms

Problem Statement
While deep learning models can detect cardiac diseases from 12-lead electrocardiograms (ECGs) with accuracy rivaling experts, their clinical adoption is hindered by a lack of interpretability. Existing Explainable AI (XAI) techniques, such as saliency maps and Grad-CAM, identify where in an ECG a model derives information but fail to explain what specific modification would alter the prediction. This gap limits their utility for clinicians, who routinely engage in differential diagnosis by reasoning about competing hypotheses (e.g., distinguishing myocardial infarction from left ventricular hypertrophy). Furthermore, standard adversarial attacks often apply perturbations uniformly or unconstrained, failing to reflect the specific features a model deems decision-relevant. There is a need for counterfactual explanations that are minimal, clinically plausible, and explicitly guide the model to a different diagnosis by modifying the most influential features.

Methodology: InterGRACE
The paper introduces InterGRACE, a framework that generates counterfactual ECGs by guiding adversarial perturbations using Integrated Gradients (IG) attributions. The method operates on the raw 12-lead ECG waveform (specifically 9 leads: I, II, III, V1–V6) from the PTB-XL dataset.

Architecture and Training: Two independently trained classifiers (Model 1 and Model 2) were developed using an InceptionTime-style 1-D convolutional architecture to distinguish Myocardial Infarction (MI) from Normal rhythms. Both models were trained on the same data split but with different random initializations.
Adversarial Optimization: Counterfactuals are generated by optimizing an adversarial variable $x_{adv}$ $x_{a d v}$ initialized from the original input $x$ $x$ . The optimization minimizes a composite loss function:
- Targeted Classification Loss ( $L_{cls}$ ): Sigmoid cross-entropy to force misclassification toward the target class (e.g., flipping MI to Normal).
- Perturbation Penalty ( $L_{perturb}$ ): An $\ell_2$ norm penalty to restrict deviations from the original signal.
- Total Variation Loss ( $L_{tv}$ ): An anisotropic total variation penalty to ensure the perturbation is smooth and locally consistent.
IG-Guided Weighting: The core innovation is the integration of IG attributions into the gradient update. IG is computed along a path from a zero baseline to the adversarial input. The raw gradient of the loss is element-wise multiplied by the normalized IG attributions ( $\nabla_{weighted} = \nabla L \odot IG$ ). This steers the perturbation specifically toward the waveform regions the classifier already deems most influential.
Evaluation Protocol: Success is measured by the attack success rate (flipping the target model's prediction) and cross-model transferability (whether the perturbation flips a second, independently trained model). Qualitative analysis involves averaging waveforms across batches to inspect lead-wise modifications.

Key Results

Attack Success: IG-guided counterfactuals successfully flipped Model 1's predictions in 67.4% of cases on the held-out test set (7,072 recordings), achieving this in both MI-to-Normal and Normal-to-MI directions.
Cross-Model Transfer: When applied to Model 2, the same counterfactuals flipped predictions in 58.8% of cases. This indicates partial transferability, suggesting the perturbations capture a mixture of shared diagnostic features and model-specific idiosyncrasies.
Localization: Qualitative inspection revealed that perturbations were not uniform. Instead, changes concentrated heavily in the ST segment and T wave (the repolarization phase), while the QRS complex remained comparatively preserved. This aligns with the clinical understanding that ischemic abnormalities are primarily reflected in repolarization.
Model Agreement: The two independently trained models showed high correlation in their original predictions, establishing a consistent baseline for interpreting the attack's impact.

Significance and Claims
The paper posits that attribution-guided counterfactual generation offers a more transparent mechanism for explaining ECG-based deep learning models compared to standard saliency methods. By concentrating changes on decision-relevant features, InterGRACE produces explanations that are more aligned with clinical reasoning patterns.

However, the author maintains a modest stance regarding clinical readiness:

Partial Transferability: The fact that counterfactuals do not fully transfer to a second model suggests that a significant portion of the perturbation exploits the specific decision boundaries of the target model rather than purely model-agnostic physiological features.
Physiological Constraints: The current method does not explicitly enforce inter-lead physiological consistency (e.g., Einthoven's triangle relationships). Consequently, generated counterfactuals may contain artifacts, such as abrupt voltage jumps or lead combinations impossible in real cardiac signals.
Conclusion: The author presents InterGRACE as an initial, largely qualitative demonstration that attribution-guided perturbation is a viable mechanism for probing ECG classifiers. They explicitly state that the tool is not yet a validated clinical decision-support system. Before deployment, the method requires the integration of physiological constraints and systematic evaluation by cardiologists to ensure the generated counterfactuals are clinically plausible and meaningful.

Technical Summary: Integrated Gradient-guided Adversarial Counterfactual Explanations of 12-lead Electrocardiograms

More like this