Publication:
Improved interpretability and feature attribution for hybrid-AI ECG models by controlling metadata leakage and feature interactions
Date
2026
Journal article
Loading...
Author(s)
Journal
EXPERT SYSTEMS WITH APPLICATIONS
Abstract
Deep learning has shown strong potential for predicting atrial fibrillation (AFib) risk from electrocardiograms (ECGs). However, in hybrid-AI settings that combine an ECG latent representation with patient metadata (age and sex in this study), it can remain unclear to what extent predictions are driven by ECG latent features or by metadata. This ambiguity arises because the ECG latent representation may already encode metadata, causing attribution quantification methods to underestimate the attribution of explicitly provided metadata. A unified framework is proposed to improve interpretability by controlling metadata leakage, modelling ECG-metadata interactions explicitly, and enabling modality-aware attribution quantification. First, a Gradient-Reversed encoder (GRe) learns an ECG latent representation that is invariant to prespecified metadata, reducing metadata leakage. Second, a Multiplicative Feature-Interaction Dense (MFID) layer reintroduces metadata through gated multiplicative interactions with the ECG latent representation. Third, Feature-Interaction Gradient Interpretation (FIGI) combines ECG-level saliency visualisation using Grad-CAM++ with integrated gradients on a latent proxy model to quantify, per recording, the relative attribution of the ECG latent representation versus metadata in the final prediction. Across in-distribution and out-of-distribution evaluation, the proposed framework yields more clinically plausible attribution patterns and demonstrates that metadata can strongly influence AFib prediction, an effect that may be obscured by naive hybridisation. In the evaluated experiments, the proposed framework improved out-of-distribution (OOD) AFib prediction performance, although the magnitude of the benefit was task- and dataset-dependent.