⚛️ quantum physics

Sleep-Disordered Breathing Detection with Quantum Recurrent Neural Networks

This paper proposes and evaluates a variational classical-quantum architecture for detecting sleep-disordered breathing, demonstrating that Quantum Recurrent Neural Networks can achieve competitive performance with significantly fewer parameters than classical models, particularly in low-data regimes, despite current hardware limitations.

Original author: Alejandro Mayorga-Redondo, José Daniel Viqueira, Diego Alvarez-Estevez, Eduardo Mosqueira-Rey, Andrés Gómez

Published 2026-11-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original author: Alejandro Mayorga-Redondo, José Daniel Viqueira, Diego Alvarez-Estevez, Eduardo Mosqueira-Rey, Andrés Gómez

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 author. For technical accuracy, refer to the original paper. Read full disclaimer

Sleep is a state of rest that most people take for granted, yet for millions, it is interrupted by a silent struggle: the breathing stops or becomes dangerously shallow. This condition, known as sleep-disordered breathing, includes episodes where airflow ceases completely or drops significantly, often leading to a drop in blood oxygen levels. Detecting these events is vital for diagnosing serious health risks, but the current method is a laborious process. Doctors must manually review hours of recordings from sensors attached to a patient's body, looking for specific patterns in the flow of air and the saturation of oxygen in the blood. This task is time-consuming and prone to human error, prompting researchers to seek help from computers. For years, standard artificial intelligence has been trained to spot these patterns, but a new frontier in computing is now being tested to see if it can do the job better, or at least with far less digital weight.

This new frontier is quantum computing, a field that uses the strange rules of the subatomic world to process information. Unlike the switches in a standard computer that are either on or off, the basic units of a quantum computer, called qubits, can exist in a state of being both on and off at the same time. This allows them to explore many possibilities simultaneously. Scientists have long wondered if these machines could learn from data more efficiently than their classical counterparts, especially when dealing with complex sequences like a heartbeat or a breathing pattern. A team of researchers in Spain set out to test this idea by building a specific type of artificial intelligence called a quantum recurrent neural network. They designed this system to act like a memory, processing a sequence of breathing data one moment at a time, while keeping a record of what came before to understand the full story of a patient's night.

The researchers took real medical recordings from a large public database of sleep studies and broke them down into small, manageable chunks. They created three different sets of data to test their system under varying conditions. The first set was small, containing only 300 examples of breathing patterns, using just the airflow signal. The second set was medium-sized with 600 examples, adding a second signal that measured oxygen levels. The third set was the largest and most difficult, with 1,200 examples that included both normal breathing and a more subtle type of breathing trouble called hypopnea, which is harder to distinguish from normal sleep. For each of these scenarios, they trained their new quantum model and compared it side-by-side with a traditional computer model that used the same basic structure but ran on standard hardware.

The results revealed a story that depends heavily on how much data is available. When the researchers gave the models only a small amount of information to learn from, the quantum version shined. It correctly identified breathing problems in 83 percent of the test cases, significantly outperforming the traditional model, which only reached 69 percent. This suggests that the quantum approach is exceptionally good at finding patterns when it cannot rely on massive amounts of training data. However, as the amount of data increased, the advantage shifted. On the medium-sized dataset, both models performed similarly, with the traditional model showing slightly better precision in identifying specific events. When the researchers moved to the largest and most complex dataset, which included the tricky hypopnea events, the traditional model pulled ahead, achieving higher accuracy across the board.

What makes these findings particularly interesting is not just the scores, but the cost of achieving them. The traditional computer model required thousands of adjustable settings, known as parameters, to learn the patterns. In contrast, the quantum model achieved its results with fewer than one hundred parameters. To put this in perspective, the quantum model was doing the work of a massive digital brain using a tiny fraction of the internal machinery. This extreme efficiency suggests that quantum systems might be able to learn complex tasks without needing the huge memory and processing power that standard computers demand. However, this efficiency came with a trade-off. The quantum model took much longer to learn, requiring hundreds of training cycles to settle on a solution, whereas the traditional model often found its answer in just a few dozen cycles. The quantum system also showed more variation in its performance from one run to the next, indicating that it is still a bit less stable than the mature technology it is trying to challenge.

The study was conducted using a high-precision simulation of a quantum computer, as the actual physical machines available today are still too small and noisy to run such complex tasks reliably. The researchers acknowledge that while their quantum model showed great promise, especially in data-scarce situations, it is not yet a finished product ready for a hospital. The path forward involves refining these models to handle the noise of real quantum hardware and exploring whether they can be applied to other types of medical data. The work suggests that as quantum technology matures, it could offer a powerful, lightweight alternative for analyzing the intricate rhythms of the human body, potentially making the diagnosis of sleep disorders faster and more accessible for everyone.

Technical Summary: Sleep-Disordered Breathing Detection with Quantum Recurrent Neural Networks

Problem Statement
Sleep-disordered breathing (SDB), characterized by apnea and hypopnea events, poses significant public health risks and requires diagnosis via polysomnography (PSG). Current diagnostic workflows rely on trained specialists manually inspecting hours of multichannel physiological recordings, a process that is labor-intensive, time-consuming, and prone to inter-scorer variability. While classical Machine Learning (ML) models, particularly Recurrent Neural Networks (RNNs), have been applied to automate this analysis, they often demand large labeled datasets, substantial computational resources, and high numbers of trainable parameters. This work investigates whether Quantum Machine Learning (QML), specifically Quantum Recurrent Neural Networks (QRNNs), can offer a competitive alternative for binary classification of respiratory sleep events, potentially achieving comparable performance with greater parameter efficiency.

Methodology
The study proposes and evaluates a hybrid variational classical–quantum architecture for classifying time-series segments of respiratory signals.

Datasets: Three datasets were derived from the Sleep Heart Health Study (SHHS-2) follow-up phase, containing 60-second windows of airflow and blood oxygen saturation (SaO2) signals resampled to 2 Hz. The datasets varied in complexity:
- Small: 300 instances, airflow only, apnea events only.
- Medium: 600 instances, airflow and SaO2, apnea events only.
- Large: 1,200 instances, airflow and SaO2, including both apnea and hypopnea events.
Models:
- Classical Baseline (CRNN): A vanilla Recurrent Neural Network with stacked layers, designed to structurally mimic the quantum workflow. It employs a global average pooling layer over the final subset of timesteps to generate a binary classification output. No regularization (e.g., dropout) was used to ensure a fair comparison.
- Quantum Model (QRNN): A variational quantum circuit featuring an exchange register ( $E$ ) and a memory register ( $M$ ). Input vectors are encoded into the exchange register via data re-uploading, followed by a hardware-efficient ansatz evolution. The memory register preserves state across timesteps, while the exchange register is measured and reset. Classification is based on the probability of measuring the $|0\rangle$ state in the first qubit.
Experimental Setup: Both models underwent hyperparameter optimization via grid search using stratified 5-fold cross-validation (repeated 5 times). The best configurations were retrained on the full training set (with a 16% validation split) and evaluated on a held-out test set. Training utilized the Adam optimizer and binary cross-entropy loss. The quantum model was executed via a high-precision emulator without simulated hardware noise.

Key Results
The comparative analysis revealed that the relative performance of the models is highly dependent on data availability and task complexity:

Small Dataset (Low Data): The QRNN significantly outperformed the CRNN, achieving an accuracy of 83.00% compared to 69.00% for the classical model. The quantum model demonstrated superior generalization capabilities under limited training conditions.
Medium Dataset: Both models achieved comparable performance (CRNN: 88.50% accuracy; QRNN: 87.83% accuracy). The classical model showed higher precision, while the quantum model achieved better recall and AUC-ROC values.
Large Dataset (High Complexity): The classical model consistently outperformed the quantum implementation across all metrics (CRNN: 80.33% accuracy; QRNN: 73.67% accuracy), likely due to the increased difficulty of distinguishing hypopnea from normal breathing.
Parameter Efficiency: The most significant finding is the disparity in model complexity. The best-performing classical models required between 15,930 and 118,439 trainable parameters. In contrast, the best-performing quantum models required only 94 to 139 trainable parameters.
Training Dynamics: The quantum models required substantially more epochs to converge (ranging from ~~71 to 464 epochs) compared to classical models (~~23 to 47 epochs) and exhibited higher variability in convergence behavior.

Significance and Claims
The paper posits that despite current hardware limitations (noise, decoherence, and small qubit counts) and the fact that experiments were conducted in emulation, variational quantum recurrent models represent a promising alternative for time-series classification. The primary contribution is the demonstration that QRNNs can achieve competitive, and in low-data scenarios superior, classification performance while utilizing several orders of magnitude fewer trainable parameters than equivalent classical architectures.

The authors conclude that while classical approaches remain more robust for complex, high-data scenarios, the parameter efficiency of the quantum approach suggests that as quantum computing technologies mature, QRNNs could become increasingly attractive for biomedical applications where model capacity and memory requirements are critical constraints. The study does not claim a universal superiority of quantum methods but highlights their potential as a viable, efficient alternative in specific regimes, particularly where data is scarce.

More like this