ECG Lead-Group Masking in Deep Learning

• Lead-group masking is a technique that isolates predefined ECG lead subsets to assess their impact on deep learning model predictions.
• It applies a masking operation on ECG tensors by zeroing selected lead groups, enabling quantitative evaluation through observed AUROC changes.
• Experimental findings show that masking specific lead groups correlates with clinical markers, enhancing model interpretability and diagnostic reliability.

Lead-group masking is a methodological approach employed in physiologically grounded model interpretability, particularly within deep learning frameworks for medical time series such as electrocardiograms (ECGs). By selectively removing—“masking”—predefined groups of leads from the input tensor, the technique quantifies the importance of specific anatomical or electrophysiological regions to model predictions. In automated arrhythmia diagnosis using lead-aware spatial attention networks, lead-group masking probes the dependency of the model on diagnostically critical regions and serves as a quantitative validation of learned physiologic relevance (Sigfstead et al., 12 Jan 2026).

1. Definition and Rationale of Lead Group Partitioning

Lead-group masking hinges on partitioning standard ECG lead sets into functionally and anatomically motivated subsets. In the context of arrhythmia classification, leads are grouped as follows: right precordial (V1–V2), extended right precordial (V1–V3), lateral (I, V5, V6), limb (I, II), and all precordials (V1–V6). Partitioning is informed by established disease signatures: for instance, arrhythmogenic right ventricular cardiomyopathy (ARVC) manifests predominantly in right precordials, while long QT syndrome (LQTS) is most evident in lateral leads due to affected repolarization.

2. Mathematical Formulation and Masking Workflow

Let $X \in \mathbb{R}^{C \times T}$ be a preprocessed ECG tensor, where $C$ is the number of leads and $T$ is the time dimension. For masking a group $G \subseteq \{1, \ldots, C\}$, define the indicator vector $\mathbb{1}_G \in \{0,1\}^C$ such that $\mathbb{1}_G(c) = 1$ if lead $c \in G$ and $0$ otherwise. The masking operation sets the designated leads to zero across all samples via element-wise multiplication:

$X^{(G)}_{c,t} = m_c \cdot X_{c,t}$

where $m = 1 - \mathbb{1}_G$. The resultant masked input $C$0 is processed by the model identically to the original tensor, and inference proceeds without architectural modification.

3. Experimental Evaluation Protocol

Lead-group masking is evaluated by comparing classification performance between the original and masked inputs for multi-class and binary arrhythmia detection tasks. The performance metric employed is the macro-averaged area under the receiver operating characteristic curve (macro-AUROC), with masking-induced performance degradation quantified as:

$C$1

where $C$2 is the unmasked baseline and $C$3 is the score after group $C$4 is masked.

Quantitative Impact Table

Lead Group Masked	ARVC ΔAUROC (%)	LQTS ΔAUROC (%)	Macro ΔAUROC (%)
V1–V3 (right precordial)	4.54	2.71	3.17
I, V5, V6 (lateral)	0.20	2.60	2.07
V1–V6 (all precordials)	7.08	6.65	6.45

The observed changes are robust, with the standard deviation of macro-AUROC across repeats ranging from 0.003 to 0.037.

4. Physiologic-Layer Interpretability and Plausibility

Performance degradation due to masking specific lead groups coincides with clinical knowledge regarding disease localization:

• For ARVC, masking V1–V3 yields the largest macro-AUROC decrease, confirming the model’s focus on right-ventricular depolarization.
• For LQTS, lateral lead masking induces the most pronounced reduction, consistent with T-wave and repolarization anomalies observable chiefly in these vectors.

This alignment with established diagnostic criteria validates the physiologic plausibility of the lead-aware spatial attention mechanism.

5. Model-Attention Mechanism and its Interaction with Masking

Within the LASAN architecture, a per-lead importance vector $C$5 is learned, modulating final feature aggregation as $C$6, where $C$7 is the embedding for lead $C$8. Masking a group $C$9 results in $T$0 for all $T$1, thereby nullifying the contribution of that lead to the feature aggregate $T$2. A substantial AUROC drop upon masking indicates a large learned $T$3 and high relevance of the corresponding anatomical region to prediction, providing direct interpretability of attention weights in clinical context.

6. Methodological Significance and Utility in Model Development

Lead-group masking is a principled tool for:

• Quantitatively verifying the model’s reliance on physiologically meaningful input regions.
• Enhancing clinician trust by aligning model reasoning with established electrophysiological knowledge.
• Informing model refinement, feature engineering, and deployment strategies for automated ECG interpretation, particularly in inherited arrhythmia screening where diagnostic accuracy depends critically on spatial analysis of conduction and repolarization patterns.

7. Limitations, Robustness, and Prospects

No confidence intervals for $T$4 were formally reported, but low variance indicates stability of masking effects. The method is contingent upon accurate lead-group selection and may require adaptation in atypical lead configurations. A plausible implication is that further integration with dynamic or patient-specific grouping could augment interpretability, especially in non-standard signal presentations.

Lead-group masking, by illuminating model dependence on anatomical ECG structure, contributes an essential dimension to the interpretability and clinical trustworthiness of deep learning models in cardiovascular diagnosis (Sigfstead et al., 12 Jan 2026).