To Predict the Incidence Rate of Xerostomia after Radiotherapy by a Neural Network Using EORTC QLQ-H&N35 and EORTC QLQ-C30 Questionnaires Dataset for Head and Neck Cancer

• Main Authors: Po-Liang Chen, 陳柏亮
• Other Authors: Tsair-Fwu Lee
• Format: Others
• Language: zh-TW
• Published: 101
• Online Access: http://ndltd.ncl.edu.tw/handle/46665591657418328758

碩士 === 國立高雄應用科技大學 === 電子工程系 === 100 === Purpose: Artificial neural networks (ANN) and quality of life (QoL) questionnaires are used to predict the incidence of xerostomia for head and neck cancer (HNC) patients at 3 months after radiotherapy. Materials and methods: In this study, 208 head and neck cancer patients were enrolled, all patients completed the Chinese version EORTC QLQ-C30 and H&N35 questionnaires. Xerostomia were defined with prior to radiotherapy and at 3 months after radiotherapy. Five different data ratio of patient number were adopted for training and testing phases. Three input data models were used, i.e., Idose (dose in salivary glands), Itotal (dose in salivary glands, patient characteristics and diagnose information) and Irobust (input parameters choose by Pearson product-moment correlation). The feed-forward pattern recognition neural networks (PR) were used and trained by using leave-one-out (LOO) validation method. For the system accuracy checking, statistical diagnosis parameters and area under receiver operation characteristic curve (AUC) were adopted. Results: Two sets of data of HNSCC and NPC were consisted. In Pearson product-moment correlation coefficient (PPMCC) analysis found that the education, family history, dry mouth before treatment, age, is prognostic important factor. ANOVA analysis for the input data models found that Irobust’s with statistical significant (p-value <0.05), but Itotal and Idose’s did not (p-value >0.05). Simulation results, for HNSCC, AUC analysis, the highest of the three groups were Irobust model D (0.76), Irobust model C (0.68) and Irobust model (0.68). For NPC, the highest AUC is Idose model (0.75), followed by Irobust model C (0.70) and Idose model D (0.69). Conclusion: ANN can provide higher prediction accuracy for xerostomia severity after 3 months of radiotherapy. This could be provided reference of treatment planning to improve better quality of life of HNC patients.