CPR: Causal Physiological Representation Learning for Robust ECG Analysis under Distribution Shifts

TL;DR

Deep ECG models struggle under distribution shifts, especially Smooth Adversarial Perturbations (SAP). CPR imposes a Structural Causal Model with a Physiological Structural Prior to disentangle invariant pathological morphology $Z_c$ from non-causal noise $Z_s$, enforcing invariance via a mask-guided dual-pathway autoencoder and loss terms $\mathcal{L}_{recon}$, $\mathcal{L}_{reg}$, $\mathcal{L}_{cons}$, and $\mathcal{L}_{adv}$. On PTB-XL, CPR achieves $F1_{SAP}=0.632$, matching the certified robustness of Randomized Smoothing while enabling single-pass inference, and outperforms key clinical preprocessing baselines. Mechanistically, CPR localizes to the P-QRS-T morphology and yields interpretable representations with content clearly class-discriminative and style absorbing noise, supporting safe, real-time deployment under distribution shifts.

Abstract

Deep learning models for Electrocardiogram (ECG) diagnosis have achieved remarkable accuracy but exhibit fragility against adversarial perturbations, particularly Smooth Adversarial Perturbations (SAP) that mimic biological morphology. Existing defenses face a critical dilemma: Adversarial Training (AT) provides robustness but incurs a prohibitive computational burden, while certified methods like Randomized Smoothing (RS) introduce significant inference latency, rendering them impractical for real-time clinical monitoring. We posit that this vulnerability stems from the models' reliance on non-robust spurious correlations rather than invariant pathological features. To address this, we propose Causal Physiological Representation Learning (CPR). Unlike standard denoising approaches that operate without semantic constraints, CPR incorporates a Physiological Structural Prior within a causal disentanglement framework. By modeling ECG generation via a Structural Causal Model (SCM), CPR enforces a structural intervention that strictly separates invariant pathological morphology (P-QRS-T complex) from non-causal artifacts. Empirical results on PTB-XL demonstrate that CPR significantly outperforms standard clinical preprocessing methods. Specifically, under SAP attacks, CPR achieves an F1 score of 0.632, surpassing Median Smoothing (0.541 F1) by 9.1%. Crucially, CPR matches the certified robustness of Randomized Smoothing while maintaining single-pass inference efficiency, offering a superior trade-off between robustness, efficiency, and clinical interpretability.

Figures (5)

• Figure 1: The CPR Framework Architecture. The model implements the Structural Causal Model via a dual-pathway design. The Content Encoder ($E_c$) extracts the invariant pathological factor $Z_c$ under the strict guidance of the Physio-Mask, while the Style Encoder ($E_s$) captures the non-causal background noise $Z_s$. The Decoder ($G$) reconstructs the signal, ensuring that $Z_c$ and $Z_s$ are sufficient to generate $X$ while remaining statistically orthogonal.
• Figure 2: Latent Manifold Topology. (Left) Baseline embedding. (Middle) CPR Content Space ($Z_c$) demonstrates clear class separation. (Right) CPR Style Space ($Z_s$) remains unstructured.
• Figure 3: Mechanism Visualization via t-SNE. (Middle) CPR Content Space ($Z_c$) distributions for clean (Blue) and adversarial (Red) samples align closely, confirming invariance. (Right) Style Space ($Z_s$) absorbs the noise perturbation.
• Figure 4: Attention Map Divergence. (b) Baseline Grad-CAM highlights artifacts. (c) CPR attention is constrained to the P-QRS-T complex.
• Figure 5: Robustness Sensitivity. CPR (Green) maintains high F1 significantly longer than Baseline (Red) as $\epsilon$ increases.

Theorems & Definitions (1)

• proof