Hybrid quantum fusion network for multimodal cardiac abnormality detection

Cardiovascular disease remains a major global health burden, motivating automated analysis of complementary cardiac modalities. This study presents a Hybrid Quantum Fusion Network (HQFN) for cross-source electrocardiogram (ECG) and phonocardiogram (PCG) representation fusion. ECG images derived from the MIT-BIH Arrhythmia and PTB Diagnostic ECG datasets and PCG recordings from the PhysioNet/Computing in Cardiology Challenge 2016 dataset were processed independently. The samples were not patient-matched; therefore, the study does not claim patient-level multimodal diagnosis. Leakage-audited classical ResNet-50 encoders produced 2048-dimensional embeddings for each modality. Fixed training, validation and test source partitions were retained, feature standardisation was fitted using training data only, and cross-source pairs were generated without replacement. In the primary concordant Normal-Normal versus Abnormal-Abnormal task, ECG-only and the principal classical fusion algorithms achieved 100% accuracy, while HQFN achieved 99.63% ± 0.72%, indicating that this task was dominated by the ECG representation. Two cross-source interaction stress tests were consequently included. On the nonlinear concordance-versus-discordance task, HQFN achieved 89.00% ± 2.21% accuracy and fell to chance when either modality was masked or permuted, demonstrating use of both sources; however, multilayer perceptron fusion achieved 92.30% ± 0.76%. On the four-class pair-composition task, HQFN was unstable, achieving 53.60% ± 16.16%, whereas multilayer perceptron fusion achieved 92.90% ± 0.55%. Finite-shot and simulated depolarising-noise analyses showed modest degradation at lower shot counts and higher noise probabilities. HQFN therefore provides a feasible nonlinear cross-source fusion mechanism, although the experiments do not demonstrate a general quantum advantage over classical fusion.