⚡ electrical engineering

QiVC-Net: Quantum-Inspired Variational Convolutional Network, with Application to Biosignal Classification

This paper introduces QiVC-Net, a novel quantum-inspired variational convolutional network that utilizes a differentiable rotated ensemble mechanism to model structured uncertainty without increasing parameters, achieving state-of-the-art performance in phonocardiogram classification on benchmark biosignal datasets.

Original authors: Amin Golnari, Jamileh Yousefi, Reza Moheimani, Saeid Sanei

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Amin Golnari, Jamileh Yousefi, Reza Moheimani, Saeid Sanei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Teaching a Computer to Listen to Hearts

Imagine you are trying to teach a computer to listen to heartbeats (specifically, the sounds of a heart, called PCGs) to tell if a person is healthy or has a problem. This is tricky because heart sounds are messy. They are often covered in background noise (like a crying baby, traffic, or a doctor's voice), and every person's heart sounds slightly different.

Most computer programs that do this are like deterministic robots: they look at a sound, make a single guess, and say, "This is normal" or "This is bad." They don't really know how sure they are. If the sound is weird or noisy, the robot might still guess with 100% confidence, even if it's wrong. This is dangerous in medicine.

The authors of this paper built a new kind of computer brain called QiVC-Net. Instead of being a rigid robot, it acts more like a wise, cautious doctor who considers many possibilities before making a decision.

The Secret Sauce: "Quantum-Inspired" Dancing

The core innovation of this paper is a special layer in the computer's brain called QiVConv. To understand how it works, let's use an analogy.

1. The Old Way: Throwing Darts Blindly

Standard "probabilistic" computer models try to handle uncertainty by adding random noise to their calculations, kind of like throwing darts blindfolded at a target. They might throw a dart slightly left, slightly right, or slightly up. But these throws are unstructured—they are just random chaos. This can sometimes confuse the computer or make it unstable.

2. The New Way: The "Quantum" Dance Floor

The authors' new method, called QiRE (Quantum-Inspired Rotated Ensemble), is different. Imagine the computer's internal weights (the rules it uses to judge heartbeats) are dancers on a floor.

The Rotation: Instead of throwing random darts, the QiRE mechanism takes the whole group of dancers and rotates them together in a specific, structured way.
The Analogy: Think of a group of dancers holding hands in a circle. If you spin the whole circle, everyone moves, but they stay connected and maintain their formation. They don't fall apart.
Why it matters: This "rotation" is inspired by how particles behave in quantum physics (specifically, "unitary evolution"). It ensures that when the computer adds uncertainty to its thinking, it does so in a way that respects the shape and structure of the data. It keeps the "geometry" of the problem intact.

The Result: The computer doesn't just guess; it explores a "cloud" of possibilities that are all logically connected. This helps it understand the shape of the uncertainty, rather than just adding random static.

The Architecture: A Two-Way Street

The paper also introduces a special building block called the RFR (Reversal Fusion Residual) block.

The Problem: Heartbeats have a rhythm. They go lub-dub, lub-dub. If you only listen to the sound from start to finish, you might miss a pattern.
The Solution: The RFR block acts like a two-way street.
1. Forward Path: It listens to the heartbeat from start to finish.
2. Backward Path: It listens to the exact same heartbeat, but in reverse (finish to start).
3. Fusion: It combines these two views using a memory unit (LSTM) to understand the full story of the heartbeat.

By looking at the sound both forward and backward, the model captures the "temporal symmetry" of the heart, making it much better at spotting irregularities.

The Training: Learning from Mistakes

The computer was trained on two massive collections of heart sound recordings:

CinC (2016): A mix of adult heart sounds from various clinics.
CirCor (2022): A huge collection of children's heart sounds, which are notoriously noisy and difficult.

The model had to deal with a "class imbalance" problem. Imagine a classroom where 80% of the students are healthy and only 20% are sick. If a teacher just guesses "Healthy" for everyone, they get 80% right but miss all the sick kids. The QiVC-Net was trained with a special "loss function" (a scoring system) that punished it heavily for missing the sick kids, forcing it to learn how to spot the rare, abnormal sounds.

The Results: A High-Performing, Calm Model

When the authors tested QiVC-Net, the results were impressive:

Accuracy: It correctly identified heart conditions about 97.8% to 97.9% of the time on both datasets. This is better than many previous methods.
Robustness: Even when the researchers added artificial noise (like static on a radio) to the heart sounds, the model didn't panic. Its performance dropped only slightly, showing it is very stable.
Calibration (The "Confidence" Check): This is the most important part. The paper shows that the model is well-calibrated.
- Analogy: If a standard model says "I am 90% sure this is a disease," it might actually only be right 70% of the time. It is overconfident.
- QiVC-Net: If it says "I am 90% sure," it is actually right about 90% of the time. It knows when it is unsure. This is crucial for medical safety.

Summary of Claims

The paper claims that by using this quantum-inspired rotation (QiRE) and the two-way listening (RFR) architecture, they created a system that:

Is highly accurate at detecting heart murmurs.
Handles noisy, real-world data better than older models.
Provides reliable "confidence scores" so doctors know when to trust the computer's answer.
Does all this without needing a massive amount of extra computing power or adding too many extra parameters.

The authors emphasize that this is a probabilistic approach, meaning it models uncertainty directly into the math of the convolution, rather than just guessing and hoping for the best.

Technical Summary: QiVC-Net: Quantum-Inspired Variational Convolutional Network

Problem Statement

Probabilistic deep learning is essential for improving model reliability and robustness in safety-critical domains like biomedical signal analysis. However, existing Bayesian deep learning frameworks often rely on simplistic isotropic noise models or mean-field assumptions that fail to capture the structured, geometry-aware variability inherent in high-dimensional parameter spaces. This limitation leads to poorly calibrated uncertainty estimates, particularly under distributional shifts common in real-world biosignals.

Specifically, in Phonocardiogram (PCG) classification, challenges include low signal-to-noise ratios, high inter-patient variability, environmental artifacts, and significant class imbalance between normal and abnormal heart sounds. While deep learning models (CNNs, RNNs, Transformers) have improved performance, they often lack robust uncertainty quantification and interpretability, which are crucial for clinical decision support.

Methodology

The paper introduces QiVC-Net, a learning framework that integrates principles of probabilistic inference, variational optimization, and geometry-preserving operations inspired by quantum transformations. The core innovation is the Quantum-Inspired Rotated Ensemble (QiRE) mechanism, instantiated within a Quantum-Inspired Variational Convolution (QiVConv) layer.

Core Components

QiVConv Layer (Quantum-Inspired Variational Convolution):
- Mechanism: Instead of injecting unstructured Gaussian noise into kernel weights, QiVConv performs differentiable, low-dimensional subspace rotations of the convolutional weights.
- Process:
  - The convolutional kernel weights are vectorized into a single global vector.
  - A random subspace is generated via QR decomposition of a Gaussian matrix.
  - Noise is sampled, normalized to a unit hypersphere (analogous to a quantum state), and subjected to a structured unitary rotation (Haar measure) within this subspace.
  - The rotated noise is combined with the orthogonal component of the original noise (subspace swap) to create a final structured perturbation.
  - Optional depolarizing noise simulates decoherence effects.
- Advantages: This approach preserves the intrinsic geometry and norm of the weight space, enabling structured uncertainty modeling without introducing additional learnable parameters beyond the mean and standard deviation. It avoids the instability of isotropic perturbations.
Reversal Fusion Residual (RFR) Block:
- To address temporal coherence in PCG signals, the network incorporates RFR blocks.
- These blocks process input through dual paths: a forward path and a time-reversed backward path, both utilizing QiVConv layers.
- Outputs are fused using LSTM layers and residual connections (concatenation followed by refinement) to capture bidirectional temporal dependencies and physiological dynamics.
Training Strategy:
- The model is trained using a composite loss function combining Categorical Cross-Entropy (CCE) and Dice Loss, with dynamic weighting to handle class imbalance.
- Variational inference is regularized via Kullback–Leibler (KL) divergence against a unit Gaussian prior.
- The framework operates on real-valued vector spaces using QR decompositions and orthogonal Haar rotations, avoiding the computational overhead of complex-valued arithmetic while mimicking unitary evolution.

Key Contributions

The authors summarize their contributions as follows:

QiVConv Layer: A probabilistic convolutional layer introducing norm-preserving subspace rotations that enhance uncertainty calibration while maintaining expressive kernel structure.
QiRE Sampling: A structured stochastic perturbation mechanism performing differentiable unitary-inspired rotations of weights, enabling Bayesian-like inference without complex-valued arithmetic.
Uncertainty-Aware Dual-Path Architecture: Integration of QiVConv with RFR blocks to capture bidirectional temporal dynamics in PCG signals without adding extra learnable parameters.
Clinically Aligned Design: A computationally efficient framework unifying uncertainty quantification, robustness to signal variation, and temporal coherence for reliable PCG-based decision support.

Experimental Results

The framework was evaluated on two benchmark datasets: PhysioNet CinC 2016 (adult, heterogeneous conditions) and PhysioNet CirCor DigiScope 2022 (pediatric, multi-site recordings).

Performance: QiVC-Net achieved state-of-the-art performance on both datasets.
- CinC: 97.84% accuracy, 98.14% specificity, 96.89% sensitivity, and 95.60% F1-score.
- CirCor: 97.89% accuracy, 99.27% specificity, 92.34% sensitivity, and 95.55% F1-score.
Robustness: Under varying Signal-to-Noise Ratios (SNR), the model maintained high discriminative capability (AUC) and accuracy, showing graceful degradation as noise increased.
Calibration: Reliability diagrams indicated well-calibrated predictions with low Expected Calibration Error (ECE), demonstrating conservative and uncertainty-aware behavior suitable for ambiguous signals.
Ablation Studies:
- Vs. Deterministic: All variational methods outperformed the deterministic baseline.
- Vs. Standard Variational: The structured QiRE mechanism consistently outperformed standard mean-field Gaussian noise (Gaussian variant) and established Bayesian baselines (Flipout, Reparameterization) in accuracy and F1-score.
- Subspace Dimension: Sensitivity analysis showed stable performance for subspace dimensions $k \in [3, 9]$ , with $k=5$ identified as the optimal balance between accuracy and computational cost.

Significance and Claims

The paper claims that QiVC-Net offers a versatile framework for advancing uncertainty-aware modeling in real-world biomedical signal analysis. Its significance lies in:

Geometry-Preserving Uncertainty: By drawing a mathematical analogy from unitary evolution, the approach enables structured uncertainty modeling that respects the intrinsic geometry of the parameter space, leading to better-calibrated predictions than isotropic noise models.
Efficiency: The method achieves these gains without adding significant computational burden or extra learnable parameters, making it suitable for lightweight deployment.
Clinical Relevance: The framework addresses critical challenges in PCG analysis, such as noise, class imbalance, and the need for reliable confidence estimates, providing a robust tool for cardiovascular disease detection.
Generalizability: The consistent performance across diverse datasets (adult vs. pediatric, controlled vs. noisy environments) highlights the framework's potential for broader application in biosignal analysis.

The authors note that while the method is conceptually motivated by quantum mechanics, it operates entirely on classical hardware using real-valued operations, bridging the gap between theoretical quantum-inspired advantages and practical deep learning implementation.