Hybrid Quantum Approaches for Multimodal Cardiovascular Disease Detection
Melvin Yashnil Ramkhelawan
Cardiovascular diseases are currently the biggest cause of global death. A high portion of global mortality and expenditure on health are caused by such diseases, therefore, detecting abnormalities earlier and more correctly in cardiac conditions is essential for timely treatment. Diagnosis of the condition in communities or at clinical screening facilities uses techniques of electrocardiogram (ECG) and phonocardiogram (PCG). Nevertheless, actual patient assessment is hampered by the brief events, inter-patient variation, high levels of noise and erratic observation conditions. Today, deep learning techniques are improving automated screening of patients with potential cardiovascular conditions, but these methods achieve improved recognition results at the cost of larger model architecture, much greater processing load and significant susceptibility to variations in the quality and recording conditions of medical sensor recordings. Given these trade-offs, this research presents a set of quantum hybrid algorithms that use variational quantum circuits to carry out dense yet descriptive features for enhancing discriminability and stability in diagnosing cardiovascular disease. The research study focuses on three main challenges of detecting cardiovascular disease. First, arrhythmic events are often intermittent and subtle, making prompt ECG interpretation difficult, especially when input quality or resolution varies across devices and transmission conditions. Second, PCG recordings are vulnerable to noise and variability during recording, which can hinder generalisation in real-world screening settings. Third, ECG and PCG provide complementary electrical and mechanical insights into cardiac function. However, many automated methods analyse them separately or use traditional fusion strategies that do not fully leverage cross-modal interactions. To address these issues, the study presents three modular hybrid quantum algorithms that follow a consistent design philosophy and reusable components, facilitating the transition from unimodal classification to multimodal decision-making for cardiovascular disease detection. The first hybrid quantum algorithm for arrhythmia detection classifies multiple arrhythmia types by analysing their corresponding images. ECG readings are input to a pre-trained ResNet50 encoding model and are passed to the quantum classifier with a 10-qubit variational quantum circuit, simulated by PennyLane’s lightning.kokkos. Each image embedding then passes through a quantum convolutional layer before the final classification output is derived. Resolution of the input ECG signal is treated as an adjustable hyperparameter rather than a fixed input scale. The algorithm was tested on 123,998 images at 224x224 size drawn from the MIT-BIH Arrhythmia and PTB datasets. Performance was evaluated by setting test set resolutions between 80px and 320px. The algorithm consistently performed well at each resolution, with negligible calibration deviation and peak accuracy around 160px, offering practical trade-offs for its implementation. The algorithm reached a 99.98% prediction accuracy within 4.5 training hours. The integration of a variational quantum layer with a robust vision backbone shows that the algorithm is capable of performing direct ECG image classification and ensuring consistent performance across multiple resolutions, effectively rendering resolution as a flexible variable controllable by the end-user, not just a static fixed dimension. The second hybrid quantum algorithm classifies PCG signals for cardiovascular disease screening. PCG recordings derived from the PhysioNet/CinC 2016 dataset were transformed into Mel-spectrograms and fed into a ResNet50 feature extraction architecture linked with a 10-qubit variational quantum circuit executed using lightning.kokkos. In addition, Gaussian noise was added to verify the robustness of the presented algorithm against different levels of acoustic clutter typically encountered in a real clinical setting. This methodology resulted in significant outcomes, achieving 99.34% accuracy, 99.33% sensitivity and 99.35% specificity for clear PCG signals. The achieved performance was comparable with, or higher than, several previously reported PCG classification studies, although direct comparison is limited by differences in datasets and evaluation protocols.. In the presence of noisy PCG data, the algorithm also exhibited robust performance, reaching 96.68% accuracy at 20 dB SNR and 80.73% accuracy at 10 dB SNR, indicating tolerance to mild and moderate synthetic noise. This hybrid quantum machine learning approach innovates by combining a variational quantum circuit with spectrogram analysis and evaluating the tolerance of the algorithm to noise, suggesting its potential to maintain high accuracy despite realistic environmental interference. The third algorithm uses a Hybrid Quantum Fusion Network (HQFN) to combine ECG and PCG classifier-output representations through a variational quantum fusion layer. The ECG inputs were six-dimensional output-logit vectors derived from the MIT-BIH Arrhythmia and PTB Diagnostic ECG datasets, while the PCG inputs were two-dimensional outputlogit vectors derived from the PhysioNet/CinC 2016 dataset. Because the ECG and PCG data were obtained independently and were neither synchronised nor patient-matched, the experiment was framed as proof-of-concept class-consistent output-level fusion rather than patient-level multimodal diagnosis. Normal ECG outputs were paired with Normal PCG outputs, while Abnormal ECG outputs were paired with Abnormal PCG outputs. The source representations were split before pairing, and SMOTE was applied only to the training representation space. The combined eight-dimensional input was processed by an eight-qubit variational quantum circuit and mapped to a binary Normal/Abnormal decision through a classical output layer. On the held-out test set, the quantum fusion algorithm achieved 99.50% accuracy, with macro-averaged precision, recall and F1-score also at 99.50%, specificity of 99.01% and ROC–AUC of 0.999994. However, a compact classical fusion baseline achieved 100% accuracy under the same experimental conditions. Therefore, the results do not demonstrate a quantum performance advantage. Instead, they establish the feasibility of applying a variational quantum circuit to compact ECG and PCG classifier outputs within a controlled output-level fusion framework. Future work should evaluate patient-matched and synchronised recordings, richer intermediate representations, discordant modality pairs and parameter-matched classical baselines. This study presents three scalable modular hybrid quantum algorithms tailored for three methods: the use of ECG only, the use of PCG only and a hybrid method that combines both ECG and PCG signals to improve the accuracy of the detection of cardiovascular diseases. Results show that quantum machine learning could potentially generate highly accurate values about the classification of both arrhythmias and heart sounds, even with issues that commonly exist in real-world scenarios, such as lower resolution inputs and added noise. Moreover, the presented algorithms are intended to function as versatile reusable blocks capable of expanding from single modality detection up to the fusion of different cardiovascular signals through quantum computation, paving the way for the utilisation of quantum machine learning in future cardiovascular disease screening.
•Jul 28, 2026•0 citations