Design Space Exploration of Hybrid Quantum Neural Networks for Chronic Kidney Disease

Muhammad Kashif 1 2 , Hanzalah Mohamed Siraj 1 2 , Nouhaila Innan 1 2 , Alberto Marchisio 1 2 , Muhammad Shafique 1 2 · 2026

Read Design Space Exploration of Hybrid Quantum Neural Networks for Chronic Kidney Disease by Muhammad Kashif 1 2 , Hanzalah Mohamed Siraj 1 2 , Nouhaila Innan 1 2 , Alberto Marchisio 1 2 , Muhammad Shafique 1 2 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-13608.txt.

Introduction

Kashif and four colleagues explore designs of hybrid quantum neural networks for diagnosing chronic kidney disease. They test 625 combinations of data encoding, circuit layout, measurement choice, and shot count. All models use ten-fold cross-validation on a curated clinical dataset. Compact circuits with suitable encodings give the best balance of accuracy, robustness, and efficiency. One strong pairing is IQP encoding with ring entanglement, and it needs no large parameter count.

Catalog record

CREATORMuhammad Kashif 1 2 , Hanzalah Mohamed Siraj 1 2 , Nouhaila Innan 1 2 , Alberto Marchisio 1 2 , Muhammad Shafique 1 2
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38219
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.13608

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Design Space Exploration of Hybrid Quantum Neural Networks for Chronic Kidney Disease Muhammad Kashif 1 2 , Hanzalah Mohamed Siraj 1 2 , Nouhaila Innan 1 2 , Alberto Marchisio 1 2 , Muhammad Shafique 1 2 Abstract Hybrid Quantum Neural Networks (HQNNs) have recently emerged as a promising paradigm for near-term quantum machine learning. However, their practical performance strongly depends on design choices such as classical-to-quantum data encoding, quantum circuit architecture, measurement strategy and shots. In this paper, we present a comprehensive design space exploration of HQNNs for Chronic Kidney Disease (CKD) diagnosis. Using a carefully curated and preprocessed clinical dataset, we benchmark 625 different HQNN models obtained by combining five encoding schemes, five entanglement architectures, five measurement strategies, and five different shot settings. To ensure fair and robust evaluation, all models are trained using 10-fold stratified cross-validation and assessed on a test set using a comprehensive set of metrics, including accuracy, area under the curve (AUC), F1-score, and a composite performance score. Our results reveal strong and non-trivial interactions between encoding choices and circuit architectures, showing that high performance does not necessarily require large parameter counts or complex circuits. In particular, we find that compact architectures combined with appropriate encodings (e.g., IQP with Ring entanglement) can achieve the best trade-off between accuracy, robustness, and efficiency. Beyond absolute performance analysis, we also provide actionable insights into how different design dimensions influence learning behavior in HQNNs. Emails: {muhammadkashif, sm12152, nouhaila.innan, alberto.marchisio, muhammad.shafique}@nyu.edu I Introduction Quantum machine learning (QML) has emerged as a promising paradigm for leveraging near-term noisy intermediate-scale quantum (NISQ) devices in data-driven applications [ 39 , 17 , 42 , 15 , 14 , 8 ] . In QML, hybrid quantum–classical neural networks (HQNNs) are among the most practical and widely studied, where variational quantum circuits (VQCs) are embedded as trainable layers within classical neural network pipelines [ 28 , 13 , 26 ] . Such hybrid models aim to combine the representational power of quantum circuits with the optimization stability and scalability of classical deep learning [ 18 , 23 ] . HQNNs also suffer from the infamous barren plateaus problem [ 32 ] , where gradients vanish exponentially with size of underlying quantum circuits, however a number of strategies exist to mitigate or overcome these issues [ 21 , 22 ] . Similarly, the quantum hardware noise can also significantly deteriorate the performance of HQNNs. To this end, the performance robustness of HQNNs has also been widely investigated under realistic noise scenarios, as well as ways to exploit noise such that it enhances the training of HQNNs [ 25 , 1 , 27 , 2 ] Despite the rapid growth of HQNN development, their design is still largely based on heuristics. Existing studies typically fix one or two components, such as the encoding strategy or the VQC structure and report results on small benchmarks [ 43 , 24 ] . However, an HQNN consists of multiple design stages that can significantly affect the overall performance of the model, including classical to quantum data encoding, structure of VQC, type of qubit measurements, and the number of shots [ 17 , 12 , 9 ] . These choices define a high-dimensional design space that strongly influences both performance and trainability, especially under NISQ constraints [ 40 , 10 , 20 , 31 ] . However, systematic design space exploration (DSE) of HQNNs for a given task remains largely unexplored. In this paper, we consider Chronic Kidney Disease (CKD) classification, a clinically relevant medical diagnosis task involving heterogeneous, partially missing, and imbalanced data. Such datasets pose challenges not only for quantum models, but also for classical networks, making them an ideal test case for realistic hybrid quantum-classical pipelines. Specifically, we conduct a systematic design space exploration framework for HQNNs for CKD classification by evaluating a large family of HQNN architectures instead of focusing on a single hand-crafted model. Fig. 1 : Motivational Analysis highlighting the need for Design Space Exploration of HQNNs. Only by changing encoding type or quantum circuit architecture (mainly differs in entanglement topology), the accuracy can vary between 30% to 80%. I-A Motivational Analysis Fig. 1 illustrates the classification accuracy obtained by HQNNs under different data encodings (left) and quantum circuit architectures (right), while keeping the dataset, training procedure, and optimization settings fixed. We observe high performance variability purely due to design choices. Depending on the selected encoding or quantum circuit design, the achieved accuracy can range from near-random performance (30–40%) to relatively strong performance (75–80%). More importantly, these variations are not caused by changes in data, training budget, or optimizer, but purely by the architectural and encoding choices. Even among reasonable and commonly used encodings (e.g., IQP, amplitude, angle, basis) and quantum circuit architectures (e.g., ring, alternating, star, strong entangling), the spread in performance remains substantial. This demonstrates that HQNN performance is extremely sensitive to design decisions and that any arbitrary choices can easily lead to significantly suboptimal models. These results highlight a critical challenge in HQNNs, i.e., there is no single universally good encoding or ansatz, and the performance landscape is highly non-trivial. Consequently, manual trial-and-error or heuristic architecture selection is unreliable and does not scale. This observation strongly motivates the need for systematic design space exploration for HQNNs, where quantum encoding, VQC structure, and model configurations must be treated as the primary optimization variables rather than fixed design choices. I-B Our Contributions • Systematic HQNN Design Space Exploration. We present a large-scale structured exploration of the HQNN design space, covering variations in data encoding, quantum circuit architectures, measurement strategies, and shot budgets within a unified framework ( Section III ). • Unified Benchmarking Pipeline. We introduce a standardized evaluation protocol based on stratified cross-validation, consistent data preprocessing, and controlled optimization settings, which enables fair and statistically robust comparison across all hybrid model configurations. • Comprehensive Real-World Case Study on CKD. We provide the first extensive benchmark of HQNNs on a clinically relevant Chronic Kidney Disease dataset, moving beyond synthetic or toy problems, and identify practical performance and stability characteristics. • Multi-Metric Performance Analysis. We evaluate models using a composite of clinically meaningful metrics and analyze the results to identify trade-offs between performance, robustness, and architectural complexity ( Section IV ). • Empirically Derived Design Guidelines for HQNNs. Based on global trend analysis across the design space, we extract actionable insights and practical recommendations for designing HQNNs under NISQ constraints ( Section V ). II Background II-A t-distributed Stochastic Neighbor Embedding (t-SNE) t-SNE is a non-linear dimensionality reduction technique for visualizing high-dimensional data in low-dimensional spaces [ 3 ] . It preserves local neighborhood structures by modeling pairwise similarities and uses a Student t-distribution to create a heavy-tailed distribution in the low-dimensional space. We use t-SNE as a validation tool to ensure multi-dataset comparability. II-B Hybrid Quantum Neural Networks (HQNNs) HQNNs combine classical and quantum components, making them suitable for NISQ devices [ 19 ] . The architecture consists of: (1) classical preprocessing, (2) classical-to-quantum data encoding layer, (3) VQC, (4) measurement layer, and (5) classical post-processing, as shown in Fig. 2 . Below, we discuss all the primary quantum components of HQNNs: Fig. 2 : HQNN design flow, from input to output. Classical-to-Quantum

Read the complete text