🤖 machine learning

Using Explainability as a Training-Time Reliability Signal for Efficient ECG Classification

The paper introduces ERTS, an explainability-based training signal that leverages Grad-CAM focus scores to filter out unreliable samples during ECG classification, thereby improving model efficiency and macro-F1 performance while reducing computational costs.

Original authors: Veerendhra Kumar Dangeti, Xiao Gu, Ying Weng, Shreyank N Gowda

Published 2026-06-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Veerendhra Kumar Dangeti, Xiao Gu, Ying Weng, Shreyank N Gowda

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 you are a coach training a team of athletes (a computer model) to recognize different types of heartbeats from an electrocardiogram (ECG) machine.

The Problem: The Exhausting Marathon
Usually, to train these athletes, you make them run through the entire library of heartbeats over and over again. You make them study every single example, even the ones they already know perfectly. It's like making a student re-read a chapter they've memorized while they are still struggling to understand a difficult paragraph. This takes a huge amount of time, energy, and computer power.

Some coaches tried a shortcut called "Progressive Data Dropout." This is like telling the athlete: "Once you get a heartbeat right, stop studying that specific one for now." This saves time, but it has a flaw. Sometimes, an athlete gets a question "wrong" (low confidence) not because they are confused, but because the question itself is messy, blurry, or poorly written (noise). The old method kept making them study these messy questions, wasting their energy.

The Solution: ERTS (The "Focus" Coach)
The authors of this paper introduced a new method called ERTS. Think of ERTS as a coach who doesn't just look at what the athlete got wrong, but how they looked at the problem.

They use a tool called Grad-CAM, which is like a heat map that shows exactly where the athlete is looking on the heart rhythm.

Good Uncertainty: The athlete is confused, but their "gaze" (attention) is locked tightly on a specific, weird part of the heartbeat that actually matters. This is a valuable learning moment.
Bad Uncertainty: The athlete is confused, but their "gaze" is wandering all over the place, looking at random static or noise. This is a wasted effort.

How It Works in Practice

The First Filter: The coach first picks out the questions the athlete is unsure about (the "hard" ones).
The Focus Check: Before making the athlete study these hard questions, the coach checks the heat map.
- If the athlete is looking at a specific, meaningful spot, the coach says, "Great, let's study this one!"
- If the athlete's attention is scattered and blurry, the coach says, "Skip this one; it's just noise. Don't waste your time."

The Results: Smarter, Not Harder
The paper tested this on three different heart rhythm datasets using three different types of "athletes" (computer models). Here is what they found:

Better Scores: The models trained with ERTS got better at diagnosing heart issues (higher accuracy) than those trained with the old methods.
Less Work: They achieved these better scores while studying significantly fewer examples (saving about 13% to 18% of the training time).
The "Why": By filtering out the "messy" questions, the models learned faster and didn't get confused by bad data.

A Key Warning
The paper also notes that you have to be careful with how strict you are. If you filter out too many questions (being too aggressive), you might accidentally throw away the really tricky but important heartbeats that the model needs to learn. The "sweet spot" was found to be a moderate filter that removes the scattered attention but keeps the focused confusion.

In a Nutshell
This paper shows that in medical AI, quality of attention matters more than just the number of examples. By using a "focus score" to decide which confusing examples are worth studying, we can train heart-monitoring AI faster, cheaper, and more accurately, without needing to re-invent the computer architecture itself. It turns explainability from a tool used after training (to check if the AI is right) into a tool used during training (to teach the AI how to learn better).

Technical Summary: Using Explainability as a Training-Time Reliability Signal for Efficient ECG Classification

Problem Statement
Training deep neural networks for clinical time-series analysis, specifically electrocardiogram (ECG) classification, is computationally expensive and resource-intensive. While recent advances in model-level efficiency (e.g., pruning, quantization) have reduced inference costs, they do not address inefficiencies inherent in the training process itself. Standard supervised learning treats all samples as equally informative throughout optimization, leading to redundant gradient updates on samples that have already been "learned."

Existing data-centric approaches, such as Progressive Data Dropout (PDD), attempt to mitigate this by excluding samples from gradient updates once they are deemed "easy" based on model confidence. However, the authors argue that confidence is an unreliable proxy for sample utility in clinical settings. In ECG data, uncertainty can arise from two distinct sources: (1) informative uncertainty, where a sample contains subtle, clinically meaningful patterns that the model has not yet learned, and (2) unreliable uncertainty, where difficulty stems from noise, acquisition artifacts, or ambiguous labels. Confidence-based selection cannot distinguish between these cases, potentially discarding valuable diagnostic examples while retaining noisy ones.

Methodology: ERTS
The paper proposes ERTS (Explainability-based Reliability Training Signal), a framework that integrates explainable AI (xAI) directly into the training loop to refine data selection. ERTS operates as a two-stage filtering mechanism built upon Progressive Data Dropout:

Candidate Selection (PDD Stage): At each epoch, a subset of uncertain samples is identified using standard PDD criteria (e.g., low prediction confidence).
Reliability Filtering (Explainability Stage): For these candidate samples, the model generates Grad-CAM attention maps to visualize which temporal regions and lead interactions drive the prediction.
- Focus Score: A metric is derived from the Grad-CAM map to quantify the "focus" of the explanation. This score measures how concentrated the model's attention is on specific, localized regions versus being diffuse across the signal.
- Filtering Logic: Samples with low focus scores (indicating diffuse, unstable, or noise-driven attention) are filtered out. Samples with high focus scores (indicating coherent, localized attention) are retained for gradient updates, even if they remain uncertain.

This approach ensures that only samples where uncertainty is supported by meaningful, clinically plausible evidence continue to influence the optimization process. The method introduces minimal computational overhead, as Grad-CAM is computed only for the candidate subset, not the entire batch.

Key Contributions

Novel Training Signal: The paper introduces the concept of using explanation quality (specifically Grad-CAM focus) as a real-time reliability signal during training, moving xAI beyond its traditional post-hoc interpretability role.
Differentiation of Uncertainty: The framework explicitly addresses the limitation of confidence-based selection by distinguishing between informative uncertainty (difficult but meaningful) and unreliable uncertainty (noisy or ambiguous).
Generalizability: The method is designed to be architecture-agnostic and compatible with various data selection strategies (confidence-based, random, or schedule-based).

Experimental Results
The authors evaluated ERTS across three public ECG datasets (PTB-XL, CPSC 2018, and Georgia 2020) and three backbone architectures (EfficientNetV2-S, ResNet-18, and MobileNetV2).

Performance vs. Efficiency: Across all dataset-architecture combinations, ERTS-enhanced variants consistently achieved higher macro-F1 scores compared to their standard PDD counterparts while requiring fewer Effective Epochs (EE). For instance, on PTB-XL with EfficientNetV2-S, ERTS improved macro-F1 by +0.0098 while reducing EE by 13.3%.
Robustness: The improvements held true regardless of the underlying selection strategy (DBPD, SMRD, or SRD), suggesting that the focus score provides information orthogonal to model confidence.
Sample Analysis: Qualitative and quantitative analysis of retained vs. filtered samples revealed that ERTS preferentially removes normal recordings with diffuse attention (likely noise) while preserving diagnostically relevant hard cases (e.g., MI, STTC) that exhibit focused attention patterns.
Threshold Sensitivity: The study found that aggressive filtering (low focus thresholds) often degraded performance by removing informative hard examples. Moderate thresholds (e.g., $\phi=0.7$ or $0.9$) provided the optimal trade-off, acting as a selective filter rather than a strict pruning mechanism.

Significance and Claims
The paper claims that explanation quality can serve as a practical, training-time signal to improve both the efficiency and reliability of clinical time-series learning. By filtering out samples where the model's uncertainty is not grounded in coherent signal, ERTS reduces redundant computation without sacrificing predictive performance.

The authors emphasize that this work extends the utility of explainability from a passive, post-hoc validation tool to an active component of the optimization process. They position ERTS as a solution for healthcare environments where computational resources are limited and where distinguishing between "hard" learning cases and "noisy" data is critical for model trustworthiness. The paper concludes that integrating interpretability into the training loop offers a promising direction for developing data-efficient and reliable machine learning systems in healthcare, though it acknowledges that future work is needed to explore alternative explanation methods and adaptive thresholding strategies.

More like this