Dual warm-start fusion versus attention-based fusion in low-label ECG-PCG classification: a controlled ablation study

Objective.Multimodal electrocardiography-phonocardiography (ECG-PCG) learning is promising for cardiac screening, but it remains unclear whether its gains are driven mainly by complex fusion architectures or by well-initialized modality-specific encoders. This study examined the relative contribution of warm-start initialization and fusion complexity in low-label ECG-PCG classification.Approach. A controlled ablation study was performed using synchronized recordings from the PhysioNet/CinC 2016 Training-A subset. ECG-only, PCG-only, cold-start fusion, single warm-start fusion, dual warm-start fusion, attention fusion, and gated-attention fusion models were evaluated across 5%, 10%, 25%, 50 %, and 100% paired-label ratios under matched training protocols. Robustness was further assessed using alternative PCG warm-start sources.Main results. Dual warm-start fusion numerically outperformed cold-start fusion across all label budgets, with AUROC gains up to +0.050 and statistically significant improvements at four of five label ratios. In contrast, attention and gated-attention fusion provided no statistically reliable advantage over simple dual warm-start concatenation and became inferior at higher label ratios despite added complexity. Multimodal gains over the strongest single modality were label-dependent, being largest at intermediate label ratios and negligible at full supervision. Additional frozen-encoder and pre-classifier feature-space analyses showed that end-to-end fine-tuning was preferable to encoder freezing, and that attention altered latent geometry without translating into superior diagnostic discrimination.Significance. The findings indicate that, in label-limited ECG-PCG classification, modality-specific encoder warm-starting is a more reliable performance driver than additional fusion complexity. These results support label-budget-aware evaluation of multimodal physiological signal models before adopting more complex fusion modules.