Scholarlink

<p>Quantum machine learning offers a novel computational paradigm for biomedical prediction, yet its real-world adoption remains limited by hardware noise and scalability constraints. This study investigates the feasibility, performance, and interpretability of a hybrid quantum neural network (QNN), using a diabetes dataset as a representative benchmark rather than a clinical endpoint. A variational QNN classifier was developed using eight structured predictors from the Pima diabetes dataset. Performance was evaluated on a held-out test set and compared with classical machine learning models, including Logistic Regression, Support Vector Machines, Random Forest, and a Shallow Multilayer Perceptron. Model interpretability was assessed using three complementary strategies. To reflect realistic deployment conditions, noisy execution was simulated, and error-mitigation techniques such as zero-noise extrapolation (ZNE) and readout-error mitigation (REM) were applied. The QNN achieved a test accuracy of 0.732, F1 score of 0.581, and ROC-AUC of 0.833. In stratified cross-validation, the QNN demonstrated the highest mean ROC-AUC (0.840 ± 0.029) and competitive accuracy (0.780 ± 0.022) compared with classical baselines. Feature analyses consistently identified glucose, diabetes-pedigree function, age, and BMI as the most influential predictors. Noisy execution reduced performance (accuracy 0.750 ± 0.021; ROC-AUC 0.820 ± 0.027), but ZNE and REM partially restored accuracy and decision-margin stability. Hybrid QNNs show interpretable learning behavior in simulated quantum execution settings, with observable sensitivity to hardware noise. Ongoing advances in noise-mitigation techniques and hardware scalability may improve the practical reliability of variational quantum models in future quantum machine learning research.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reliability assessment of variational quantum neural networks under noise models and error-mitigation strategies in the NISQ regime

  • 期刊论文

  • 发表日期：

    2026年2月5日

查看全文
  • Kwang Hyeon Kim,
  • Jae-Keun Cho

摘要

Quantum machine learning offers a novel computational paradigm for biomedical prediction, yet its real-world adoption remains limited by hardware noise and scalability constraints. This study investigates the feasibility, performance, and interpretability of a hybrid quantum neural network (QNN), using a diabetes dataset as a representative benchmark rather than a clinical endpoint. A variational QNN classifier was developed using eight structured predictors from the Pima diabetes dataset. Performance was evaluated on a held-out test set and compared with classical machine learning models, including Logistic Regression, Support Vector Machines, Random Forest, and a Shallow Multilayer Perceptron. Model interpretability was assessed using three complementary strategies. To reflect realistic deployment conditions, noisy execution was simulated, and error-mitigation techniques such as zero-noise extrapolation (ZNE) and readout-error mitigation (REM) were applied. The QNN achieved a test accuracy of 0.732, F1 score of 0.581, and ROC-AUC of 0.833. In stratified cross-validation, the QNN demonstrated the highest mean ROC-AUC (0.840 ± 0.029) and competitive accuracy (0.780 ± 0.022) compared with classical baselines. Feature analyses consistently identified glucose, diabetes-pedigree function, age, and BMI as the most influential predictors. Noisy execution reduced performance (accuracy 0.750 ± 0.021; ROC-AUC 0.820 ± 0.027), but ZNE and REM partially restored accuracy and decision-margin stability. Hybrid QNNs show interpretable learning behavior in simulated quantum execution settings, with observable sensitivity to hardware noise. Ongoing advances in noise-mitigation techniques and hardware scalability may improve the practical reliability of variational quantum models in future quantum machine learning research.