Pretreatment Radiation Esophagitis Prediction Using Quantum Machine Learning In Esophageal Cancer Patients
Abstract
Purpose
To introduce a hybrid quantum-classical machine learning (QML) approach and validate its feasibility and accuracy for pretreatment radiation esophagitis (RE) prediction in patients with esophageal cancer (EC) undergoing radiotherapy and/or chemotherapy.
Methods
This study enrolled 218 EC patients (training and internal validation) from hospital one and 55 EC patients (external validation) from hospital two, with grade ≥ 2 RE incidences of 64 and 20, respectively. Dose distribution images were converted into quantum states via angle encoding. Quantum features (Q) were extracted using three quantum models: 1) classical convolutional neural network (CNN) with quantum convolution (Q-CNN), 2) Q-CNN plus classical Attention (Q-CNN+Attention), and 3) Q-CNN plus quantum Attention (Q-CNN+Q-Attention). The hybrid QML model integrates handcrafted dosiomic features (D), Quantum features (Q), and clinical factors (C) through a feature-level concatenation. The concatenated feature vector was processed by a Random Forest classifier for RE prediction.
Results
Models using only quantum features achieved an accuracy of 0.70 (Q-CNN), 0.83 (Q-CNN +Attention), and 0.80 (Q-CNN+ Q-Attention) in external validation, respectively. Feature fusion (Q+D+C) improved the accuracy of models in comparison with quantum features alone. Q (Q-CNN+Q-Attention) + D + C demonstrated optimal performance, achieving accuracy, sensitivity, and specificity of 0.85, 0.83, 0.92 (training); 0.80, 0.73, 0.84 (internal validation); and 0.83, 0.73, 0.89 (external validation), respectively. The Random Forest model using fused features Q+D+C extracted via Q-CNN+Q-Attention achieved AUCs of 0.89 (training), 0.89 (internal validation), and 0.83 (external validation), outperforming models using only Q-CNN+Q-Attention (AUCs: 0.78 training, 0.78 internal, 0.80 external).
Conclusion
This study proposes a novel QML method for pretreatment RE prediction. By integrating quantum amplitude encoding, quantum attention mechanisms, and multimodal feature fusion (Q+D+C), the model enhances prediction accuracy and reliability, demonstrating significant potential for clinical application.