Predictive Performances Analysis of Statistical Learning Methods on Retrospective Multivariate Normal Tissue Complication Probability Models Investigating Causing Xerostomia after Head and Neck Cancer Radiotherapy

• Main Authors: LIOU,MING-HSIANG, 劉明祥
• Other Authors: LEE,HSIAO-YI, LEE,TSAIR-FWU
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/31298312979684727567

博士 === 國立高雄應用科技大學 === 電機工程系博碩士班 === 104 === Purpose: The factors influencing xerostomia complication probability of head and neck cancer (HNC) patients treated by radiotherapy are unknown and complicated. Current analysis models are too simple and are not accurate. The present study uses different statistical learning methods to obtain the clinical predictive factors, and increases the predictive performance of multivariate normal tissue complication probability (NTCP) models. Materials and Methods: This is a retrospective study investigating the treatment of HNC using two different modalities, linear accelerator intensity-modulated radiotherapy (IMRT) and helical tomotherapy (HT). Data were collected for 122 head and neck squamous cell carcinoma patients treated by IMRT (HNSCC_ IMRT), 84 nasopharyngeal carcinoma patients treated by IMRT (NPC_ IMRT), and 67 NPC treated by HT (NPC_HT). Quality of life of the cancer patients was evaluated using the European Organization for Research and Treatment Cancer (EORTC) questionnaires: QLQ-H&N35 and QLQ-C30 (traditional Chinese versions). In the xerostomia causing factors assessment, two models were used to discuss the relationship between dose and clinical factors. The first model used parotid mean doses to create the Lyman-Kutcher-Burman (LKB) NTCP model of general radiotherapy. The second model used three stastistical learning methods were used to select multivariate predictive factors using traditional statistics, forward stepwise selection and the least absolute shrinkage and selection operator (LASSO). The logistic regression was then used to create the xerostomia predictive model. Finally, according models optimization test and models predict power test to evaluate the performance of this system. Results: The predictive influencing factors of mild-to-severity xerostomia 1-3 months after radiotherapy (acute effects) were selected with LASSO. The factors were the mean dose to the ipsilateral parotid gland (Dip), contralateral parotid gland (Dcp), and age for HNSCC_ IMRT patients. For NPC_IMRT patients, the factors were the Dip and Dcp, financial circumstance, and age. For NPC_HT, the factors were the mean dose to the ipsilateral submandibular gland (Dis), contralateral submandibular gland (Dcs), Dcp, and xerostomia before radiotherapy. The area under the receiver operation characteristic curve (AUC) of the three models was 0.88, 0.87 and 0.96. The predictive influencing factors of mild-to-severity xerostomia 6-12 months after radiotherapy (Late Effects) were also selected with LASSO. The factors were the Dip, Dcp, and T stage for HNSCC_IMRT patients. For NPC_IMRT patients, the factors were the Dip, Dcp, and education. For NPC_HT patients, the factors were age, mean dose to the oral cavity, education, and T stage. The AUC of the three models was 0.98, 0.96 and 0.95. Conclusions: The comparison of three statistical learning methods used in this xerostomia study suggests that the LASSO has better predictive power than either stepwise selection or traditional statistic and LKB NTCP models. Therefore, the LASSO method may be better for multivariate NTCP modeling.