PQA 10 - PQA 10 Head & Neck Cancer and Health Services Research/Global Oncology Poster Q&A
3770 - Theoretical Research on Temporal Lobe Injury after Intensity-Modulated Radiotherapy for Nasopharyngeal Carcinoma Using Machine Learning Approach and its Clinical Application
The Second Affiliated Hospital of Nanchang University Nanchang, Jiangxi
L. Zeng1, H. D. Ouyang2, and K. Zhu3; 1Department Of Oncology, The Second Affiliated Hospital of Nanchang University, Nanchang, China, 2Vocational Teachers College, Jiangxi Agricultural University, Nanchang, China, 3Department of Oncology, The Second Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, China
Purpose/Objective(s): The purpose of this study is to compare various NTCP models for predicting the incidence of radiation-induced temporal lobe injury (TLI) in patients with nasopharyngeal carcinoma (NPC) undergoing intensity-modulated radiotherapy (IMRT) using advanced machine learning methodologies. The ultimate aim is to generalize the optimized model for practical clinical application and extend the machine learning modeling approach to NTCP prediction models for other normal tissues. Materials/
Methods: From 2003 to 2018, newly diagnosed NPC patients without distant metastasis from two academic institutions (Sun Yat-sen University Cancer Center [Data-A; n = 278 (training cohort)] and [Data-C; n = 215 (internal validation cohort)] and Second Affiliated Hospital of Nanchang University [Data-B; n = 119 (external validation cohort)]) received radical IMRT treatment. A large number of dose-volume effect models were fitted for the incidence of TLI after IMRT for NPC. The following models were considered: (1) Lyman model, (2) Logit formulas of generalized equivalent uniform dose (EUD) simplified from DVH, (3) sequential rebuilding unit (RU) model, (4) Poisson-EUD model, and (5) mean dose model. Parameter estimation was performed using maximum likelihood method. R2, AIC values, and AUC values were used to compare the performance of the models. Results: In this research, a range of NTCP models were utilized, such as Lyman-EUD, Logit-EUD, Poisson-EUD, sequence reconstruction unit, and logistic-EUD, all derived from Dataset A. After conducting a comprehensive evaluation considering metrics such as AIC, R2, AUC values, and others, it was found that the logistic-EUD model constructed using D0.5cc and D1cc parameters exhibited the best fit to the data. For the logistic-EUD model with D0.5cc, the parameter "a" was determined to have a value of 25.568 (with a 68% confidence interval [CI] of 17.676~40.614), while the parameter ß0 was -16.053 ( 68% CI -21.889~-13.681), and the parameter ß1 was 0.203 ( 68% CI 0.173~0.269).Similarly, for the logistic-EUD model with D1cc, the parameter "a" was found to be 21.720(68% CI 12.981~23.361), the parameter ß0 was -17.000( 68% CI -17.750~-14.218), and the parameter ß1 was 0.215 ( 68% CI 0.178~0.226). Its crucial to recognize that only the parameter values of the logistic-EUD model fell within the 68% confidence interval of the estimated parameters in the training set. These findings were consistent in both the internal validation set and the external validation set. Conclusion: Utilizing machine learning, we have substantiated the logistic model as the utmost precise predictive model for post-IMRT incidence of TLI in NPC. Moreover, we have developed an open-source and cost-free software tool that enables the computation of radiation-induced TLI rates over a span of five years using diverse NTCP models. This invaluable resource is accessible to radiation oncologists worldwide.