Explainable quantum-enhanced machine learning for hypertension prediction

Eur. Phys. J. Spec. Top. https://doi.org/10.1140/epjs/s11734-025-01629-5 THE EUROPEAN PHYSICAL JOURNAL SPECIAL TOPICS Regular Article Explainable quantum-enhanced machine learning for hypertension prediction Turker Berk Donmez1,a 1 2 3 and Mustafa Kutlu2,3 Biomedical Engineering, Sakarya University of Applied Sciences, Sakarya, Turkey Mechatronics Engineering, Sakarya University of Applied Sciences, Sakarya, Turkey System Engineering, Military Technological College, Muscat, Oman Received 30 December 2024 / Accepted 7 April 2025 © The Author(s) 2025 Abstract Chronic disease prediction presents ongoing challenges in healthcare, primarily due to the complexity of medical data and the need for models that are both accurate and interpretable. This study introduces a quantum-enhanced machine learning model specifically designed for the prediction of hypertension, combining quantum feature transformation with classical algorithms to deliver precise and reliable results. The model demonstrates high performance, achieving an accuracy of 98.40%, precision of 99.3%, recall of 98.6%, and an F1-score of 98.9%. To ensure transparency and facilitate clinical interpretation, explainable AI (XAI) techniques are employed through SHAP values, highlighting critical factors such as hypertension drug usage, age, ferritin, and cholesterol levels as key contributors to hypertension prediction. This quantum-based approach exemplifies the potential for leveraging cutting-edge technologies in healthcare, offering a robust solution that not only ensures predictive accuracy but also supports interpretability—essential for informed clinical decision-making. The integration of quantum computing and explainable machine learning represents a promising step forward in the development of predictive models tailored to complex medical datasets. 1 Introduction Quantum computing has garnered significant attention in recent years due to its potential to revolutionize various computational tasks by harnessing the principles of quantum mechanics. The ability of quantum computers to perform complex calculations exponentially faster than classical systems makes them particularly promising in domains such as machine learning, optimization, and data analysis [1, 2]. One of the critical areas where quantum computing can have a transformative impact is healthcare, specifically in the early detection and prediction of chronic diseases. Machine learning models, such as XGBoost, have demonstrated substantial efficacy in analyzing large datasets and predicting outcomes in medical diagnostics. However, as the complexity of data increases, traditional algorithms may face limitations in scalability and interpretability [3, 4]. Recent advancements in hybrid quantum-classical models offer an innovative approach to overcoming these challenges by combining the strengths of quantum computing and classical machine learning. Quantum-enhanced machine learning integrates quantum computing’s capabilities of parallelism and entanglement with classical algorithms to improve predictive accuracy and computational efficiency [5, 6]. This approach is particularly relevant for chronic disease prediction, where large-scale datasets with numerous features and complex relationships need to be processed efficiently [7, 8]. Hybrid models, such as the XGBoost–quantum framework proposed in this paper, can significantly enhance the performance of predictive models by leveraging quantum resources to complement classical machine learning techniques. One promising advancement in quantum computing is the development of quantum encoding of 3D spatial information, which demonstrates a novel way of encoding complex data into qubits. Instead of using raw Cartesian coordinates, this approach transforms spatial information into relative position tuples that preserve key invariance and equivariance properties. For example, the Atom2Qubit strategy maps atomic features onto qubits using angle encoding, a method particularly effective for representing 3D molecular structures [9]. By leveraging parameterized a e-mail: (corresponding author) 0123456789().: V,-vol 123 Eur. Phys. J. Spec. Top. quantum circuits (PQC), researchers have shown that quantum models can efficiently encode geometric relationships, making them highly suitable for tackling complex, high-dimensional problems—such as molecular property prediction and 3D geometry generation. Notably, this method has been successfully applied to molecular datasets like QM9, demonstrating competitive performance with significantly fewer trainable parameters than classical deep learning models. The ability to embed spatial information into quantum circuits highlights the potential for quantum computing in healthcare applications, where structural and feature-based relationships in biomedical data can be encoded and processed efficiently. Building upon these quantum-based advancements, another key development is the quantum variational autoencoder for 3D molecule generation (QVAE-Mole). This fully quantum variational autoencoder is specifically designed to generate 3D molecular structures, thereby surpassing prior methods that focused solely on molecular graphs. By utilizing amplitude encoding, QVAE-Mole maps molecular spatial data onto quantum states efficiently, requiring only O(C log n) qubits—which makes it feasible for near-term quantum hardware. Unlike classical VAEs that rely on Gaussian distributions in the latent space, QVAE-Mole employs the von Mises–Fisher (vMF) distribution, a choice that naturally aligns with quantum states and preserves the unit-norm constraint [10]. This spherical latent space enhances molecular representation, allowing the quantum model to maintain geometric consistency while encoding molecular features. In addition, the quantum encoder–decoder architecture of QVAE-Mole leverages parameterized quantum circuits (PQC) for efficient learning. The encoder processes 3D molecular features using quantum state tomography and entangling gate operations, while the decoder reconstructs molecular geometries through amplitude decoding. This approach is further extended with conditional generation capabilities in the form of QVAE-Mole, enabling the targeted design of molecules based on desired chemical properties such as synthetic accessibility and molecular stability. Empirical evaluations on the QM9 dataset indicate that QVAE-Mole achieves chemically plausible molecule generation with a drastically reduced number of parameters compared to classical VAEs, demonstrating the efficiency of quantum models in molecular property prediction. Beyond advancements in quantum molecular representation, the broader field of hybrid quantum-classical machine learning (HQCLML) has emerged as a powerful strategy for tackling complex, high-dimensional computational challenges. Traditional machine learning methods often struggle with simulating intricate molecular interactions and optimizing large-scale systems due to computa (...truncated)