QiVC-Net: Quantum-Inspired Variational Convolutional Network with Rotated Ensembles for Murmur Classification

Abstract

Phonocardiogram (PCG) classification is a critical step in computer-aided auscultation for early detection of cardiovascular disorders, yet real-world recordings are often degraded by noise, variability, and limited labeled data. This study introduces a quantum-inspired variational convolutional network that integrates probabilistic reasoning, temporal symmetry, and structured uncertainty modeling for robust PCG classification. The core innovation is the quantum-inspired rotated ensemble sampling mechanism, which perturbs latent weights via differentiable low-dimensional subspace rotations inspired by quantum state evolution, preserving geometric coherence while enriching posterior expressiveness. The proposed framework employs a lightweight architecture coupled with a physiologically informed preprocessing pipeline, including band-pass filtering, spike artifact removal, multi-envelope feature extraction, and beat-synchronous segmentation. This design enhances robustness to inter-subject variability and improves uncertainty calibration without increasing parameter count. Evaluation on the PhysioNet/CinC 2016 dataset demonstrates state-of-the-art performance, achieving 97.7% classification accuracy alongside well-calibrated probabilistic predictions, underscoring its potential for reliable, uncertainty-aware clinical decision support in heart sound analysis.