Key result
Intermediate fusion outperforms late fusion in multimodal ECG classification, reaching ~97% peak accuracy.
Why the study?
The study investigated the effectiveness of intermediate versus late fusion strategies for multimodal ECG classification across time, frequency, and time-frequency domains to identify the most efficient and interpretable configuration.
Does an intermediate fusion strategy improve classification accuracy and interpretability compared to late fusion in deep neural networks for multimodal ECG signals?
Does an intermediate fusion strategy improve classification accuracy and interpretability compared to late fusion in deep neural networks for multimodal ECG signals?
Standardized Mean Difference: 0.4
Intermediate fusion outperforms late fusion in multimodal ECG classification, and the proposed Explainable-AI Trustworthiness (EAT) framework provides measurable criteria for evaluating model interpretability.
Intermediate fusion may boost ECG AI accuracy in research datasets; leaves open clinical validation and real-world interpretability gains.
This study investigates the effectiveness of intermediate and late fusion strategies for multimodal ECG classification across time, frequency, and time–frequency domains. Multiple fusion architectures were designed and evaluated to identify the most efficient and interpretable configuration. Intermediate fusion consistently outperformed late fusion, achieving a peak accuracy of 97 %, with a large effect size (Cohen’s d > 0.8) relative to single-domain models and d = 0.4 compared with late fusion. Interpretability analyses using saliency maps revealed that model attention aligned closely with physiologic ECG regions, while Mutual Information (MI) confirmed a strong statistical dependency between discretized ECG signals and their corresponding saliency distributions. We introduce the Explainable-AI Trustworthiness (EAT) theory, which defines the foundational principles of informational grounding, robustness, and architectural faithfulness as necessary conditions for trustworthy explanations in biomedical AI. Building on this theoretical foundation, we propose the EAT framework, which operationalizes these principles intoformal, measurable criteriafor ECG model interpretability. The framework quantifies (i) informational grounding to expert-annotated ST–T segments through adjusted/normalized MI (AMI/NMI); (ii) robustness to physiologic and adversarial perturbations via flip-rate, Δptrue, and saliency stability; and (iii) architectural faithfulness through branch concordance and ablation consistency. Under these criteria, EAT exhibits statistical soundness, monotonicity, and completeness, advancing explainability from subjective visualization toward falsifiable, auditable, and clinically trustworthy AI evaluation. The integration of EAT theory and framework establishes a unified foundation that bridges theoretical rigor with practical model validation, promoting reproducibility, clinical credibility, and cross-domain generalization in future biomedical AI research.
No takes yet. Share an insight, caveat, or question.
Oladunni et al. (2025) studied ECG classification. Intermediate fusion strategy vs. Late fusion strategy and single-domain models was evaluated on Peak accuracy for multimodal ECG classification (Cohen's d = 0.4). Intermediate fusion for multimodal ECG classification outperformed late fusion, achieving a peak accuracy of 97% with an effect size of Cohen's d = 0.4 compared to late fusion.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: