💻 computer science

Quantum Bayesian Neural Network for Uncertainty-Aware ECG Anomaly Detection

This study proposes a Quantum Bayesian Neural Network that integrates multi-scale feature engineering, variational quantum learning, and Bayesian inference to achieve high-performance, uncertainty-aware ECG anomaly detection, demonstrating that a six-qubit Dense-Angle encoding model with data re-uploading achieves an AUC of 0.91 while validating its robustness through execution on physical IBM quantum hardware.

Original authors: Van Hieu Nguyen, Sutharshan Rajasegarar, Shiva Raj Pokhrel

Published 2026-10-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Van Hieu Nguyen, Sutharshan Rajasegarar, Shiva Raj Pokhrel

Original paper licensed under CC BY 4.0 (https://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 human heart speaks in a language of electrical pulses, a rhythmic code that doctors have long relied upon to diagnose illness. When this rhythm falters, the resulting irregularities, known as arrhythmias, can signal serious health risks. Detecting these anomalies in a heartbeat is a task that computers have learned to perform, yet the signals are often messy and vary greatly from person to person. This complexity leaves a gap: a computer might spot a problem, but it cannot always say how sure it is. In the world of medical data, knowing the level of certainty is just as vital as the diagnosis itself. To address this, researchers are exploring a new frontier of computing that moves beyond the binary logic of traditional machines. They are turning to quantum computing, a field that uses the strange, probabilistic rules of subatomic particles to process information. By combining this powerful new hardware with a statistical approach called Bayesian learning, which treats knowledge as a matter of probability rather than absolute fact, scientists hope to build systems that not only find errors in heartbeats but also understand the confidence behind their findings.

A team of researchers at Deakin University in Australia has taken a significant step in this direction by developing a specialized system to detect heart anomalies using these quantum principles. Their work focuses on a specific type of artificial intelligence called a Quantum Bayesian Neural Network. Unlike standard computer programs that follow a fixed set of rules, this system is designed to learn from data while simultaneously tracking its own uncertainty. The researchers tested their approach using a vast collection of real heart recordings, specifically the MIT-BIH Arrhythmia Database, which contains thousands of annotated heartbeats from patients. They did not simply feed this raw data into a computer; instead, they first broke the signals down into smaller, manageable pieces and extracted a wide variety of features, such as the shape of the wave, its frequency, and how different parts of the signal relate to one another over time. This careful preparation ensured that the quantum computer received a clear and rich description of the heartbeat before the analysis began.

The core of their experiment involved mapping these heart features onto a quantum processor. The researchers tried several different ways to translate the heart data into the language of quantum bits, or qubits. They compared a method where each feature was assigned to a single qubit against a denser approach where two features were packed onto each qubit. They also tested how the depth of the quantum circuit, or the number of processing layers, affected the results, and whether feeding the data back into the system multiple times improved its learning. After running thousands of simulations, they found that the most successful configuration used the denser encoding method with six qubits and four layers of processing, while also re-introducing the data during the calculation. This specific setup achieved a high level of accuracy in distinguishing normal beats from abnormal ones, with a score of 0.91 on a scale where 1.0 is perfect. More importantly, this configuration provided the most reliable probability estimates, meaning the system's confidence levels were well-calibrated to the actual likelihood of an anomaly.

However, a simulation is not the same as reality. To see if their findings held up in the physical world, the team took their most compact and efficient model—a smaller version using four qubits—and ran it on a real quantum computer, the IBM Marrakesh quantum processor. This machine, unlike the perfect simulations, is subject to the noise and imperfections of the physical world. The results were revealing. While the real hardware did not perform quite as well as the simulation, it retained the ability to detect heart anomalies with surprising consistency. The system correctly identified the vast majority of abnormal beats, even on the noisy machine. The researchers observed that the main difference was a slight increase in false alarms, where normal beats were occasionally flagged as abnormal, but the system rarely missed a true anomaly. This demonstrated that the learned behavior of the quantum model could survive the transition from a theoretical environment to a physical one.

Perhaps the most valuable discovery of the study was how the system handled its own uncertainty. Because the model was built on Bayesian principles, it could generate a measure of doubt for every single prediction it made. The researchers found that when the quantum computer was unsure, its predictions were indeed less reliable. By using this internal measure of uncertainty, they could filter out the most doubtful results. When they did this, the accuracy of the remaining predictions jumped significantly, rising from about 63 percent to nearly 87 percent. This suggests that in a real-world medical setting, such a system could act as a smart filter, flagging only the cases where it is highly confident, or alerting a human doctor when it is uncertain. The study concludes that while quantum computers are still in their early stages, combining them with Bayesian learning creates a powerful tool for medical analysis. It offers a path forward where machines do not just give answers, but also tell us how much we should trust them, a crucial step for any technology intended to assist in human health.

Technical Summary: Quantum Bayesian Neural Network for Uncertainty-Aware ECG Anomaly Detection

Problem Statement
Automated anomaly detection in electrocardiogram (ECG) signals remains challenging due to substantial temporal, morphological, and inter-patient variability. While Quantum Machine Learning (QML) offers a framework for learning complex representations from biomedical signals, its effectiveness is contingent upon data encoding strategies, variational circuit configurations, and reliability under physical quantum noise. Furthermore, existing quantum approaches often lack mechanisms to quantify predictive uncertainty, which is critical for robust medical decision-making. Current literature largely addresses quantum ECG classification, Bayesian quantum learning, and variational circuit design independently, leaving a gap in unified frameworks that integrate these elements for time-series anomaly detection under physical execution constraints.

Methodology
The study proposes an uncertainty-aware Quantum Bayesian Neural Network (QBNN) pipeline evaluated on the MIT-BIH Arrhythmia Database. The methodology proceeds through four distinct stages:

Data Preprocessing and Feature Engineering:
- Leakage-Safe Pipeline: The dataset is partitioned into training and validation sets prior to any data-dependent preprocessing to prevent information leakage.
- Multi-Window Analysis: A record-adaptive strategy selects three optimal window lengths from nine candidates based on morphology similarity, mutual information, signal entropy, SNR, and sequence stability. Features are extracted from time, frequency, wavelet, and signal-quality domains.
- Dependency-Aware Features: The representation is augmented with features capturing linear, monotonic, nonlinear, and temporal dependencies (Pearson/Spearman correlation, mutual information, lag interactions, and graph-based relationships).
- Feature Selection: A pipeline reduces 979 engineered features to a ranked pool of 120 (60 temporal, 22 dependency, 20 frequency, 11 signal quality, 5 rhythm, 2 other) using low-variance filtering, correlation filtering, and ranking. Robust standardization is applied using training-derived statistics.
Quantum Circuit Architecture:
- Encoding Strategies: Two encoding schemes are evaluated: Angle Encoding (one feature per qubit via rotation) and Dense-Angle Encoding (two features per qubit via paired $R_Y$ and $R_Z$ rotations).
- Configurations: Experiments vary qubit counts (4 and 6), variational layers (2 and 4), and the presence of data re-uploading (repeatedly introducing input features across layers).
- Readout: Both architectures utilize $Z + ZZ$ observables to capture single-qubit and pairwise correlations.
Bayesian Inference Integration:
- Instead of deterministic point estimates, trainable quantum parameters $\theta$ are assigned Gaussian priors.
- A factorized Gaussian variational distribution approximates the posterior, with means optimized during training and standard deviations fixed for stability.
- Predictions are generated by marginalizing over posterior samples ( $S=4$ ), yielding posterior predictive anomaly probabilities.
- Training combines binary cross-entropy loss with a KL-divergence regularizer, utilizing a warm-up schedule.
Hardware Evaluation:
- A compact Dense-Angle Q4–L2 model is deployed on the IBM Marrakesh quantum processor using Qiskit Runtime.
- The experiment compares ideal statevector simulation against physical execution (4,096 shots per circuit) to assess predictive consistency and noise resilience.

Key Contributions

Systematic Architecture Evaluation: The study evaluates the interaction between encoding density, qubit width, variational depth, and data re-uploading, demonstrating that performance is architecture-dependent rather than a function of complexity alone.
Uncertainty-Aware QBNN: It integrates multi-window and dependency-aware ECG representations with Bayesian variational quantum learning, enabling posterior predictive inference and prediction-level uncertainty quantification.
Classical-Quantum Alignment: The work investigates the joint influence of input scaling, parameterization, and optimization settings on QBNN performance.
Hardware-Aware Validation: It validates the model on physical quantum hardware, extending evaluation to statevector-hardware consistency and using uncertainty metrics for error identification.

Results

Simulation Performance: Among 16 configurations, Dense-Angle encoding with 6 qubits, 4 layers, and data re-uploading achieved the strongest probabilistic performance:
- AUROC: 0.91
- AUPRC: 0.90
- Recall: 0.88
- Brier Score: 0.13
- Notably, the configuration with the highest threshold-based classification accuracy (Angle Q6-L2 without re-uploading, Accuracy 0.86) did not yield the best probabilistic discrimination, highlighting the distinction between classification and probabilistic performance.
Hardware Execution: The compact Dense-Angle Q4–L2 model deployed on IBM Marrakesh showed:
- Accuracy: 0.630; Balanced Accuracy: 0.666.
- AUROC: 0.831; AUPRC: 0.733.
- High probability-level agreement with statevector simulations (Pearson correlation $r=0.9933$ , 94.8% prediction agreement).
- Performance degradation was primarily driven by an increase in false positives, while anomaly sensitivity (recall) remained high (0.959).
Uncertainty Utility: Bayesian uncertainty metrics (predictive variance, entropy) effectively identified unreliable predictions.
- Incorrect hardware predictions exhibited significantly higher uncertainty than correct ones.
- Selective Inference: Excluding the most uncertain samples to retain 30% coverage increased hardware accuracy from 0.630 to 0.867.
- Uncertainty metrics achieved AUROC values of ~0.76 for detecting hardware-induced errors.

Significance and Claims
The paper claims to demonstrate the feasibility of uncertainty-aware quantum learning for ECG analysis. It establishes that:

Architecture Synergy: QBNN performance is governed by the interaction of encoding strategy, circuit capacity, and data re-uploading, rather than circuit complexity in isolation. Dense-Angle encoding offers a compact representation, while data re-uploading benefits are specific to certain architectures (e.g., deeper Dense-Angle models).
Probabilistic vs. Deterministic: The study highlights that the configuration yielding the highest classification accuracy is not necessarily the one providing the strongest probabilistic calibration.
Hardware Viability: Despite physical noise, the learned anomaly detection behavior persists, and Bayesian uncertainty provides a practical signal for identifying hardware-sensitive predictions and enabling selective inference.
Limitations: The authors modestly note that experimental scales are limited to balanced subsets of the MIT-BIH database and a single quantum backend. Conclusions regarding cross-dataset generalization and hardware robustness are restricted by these constraints. Future work is proposed to scale the QBNN to larger datasets and circuits and to investigate hardware error mitigation.

Technical Summary: Quantum Bayesian Neural Network for Uncertainty-Aware ECG Anomaly Detection

More like this