BACKGROUND AND PURPOSE: Design a model for prediction of acute dysphagia following intensity-modulated radiotherapy (IMRT) for head and neck cancer. Illustrate the use of the EMLasso technique for model selection. MATERIAL AND METHODS: Radiation-induced dysphagia was scored using CTCAE v.3.0 in 189 head and neck cancer patients. Clinical data (gender, age, nicotine and alcohol use, diabetes, tumor location), treatment parameters (chemotherapy, surgery involving the primary tumor, lymph node dissection, overall treatment time), dosimetric parameters (doses delivered to pharyngeal constrictor (PC) muscles and esophagus) and 19 genetic polymorphisms were used in model building. The predicting model was achieved by EMLasso, i.e. an EM algorithm to account for missing values, applied to penalized logistic regression, which allows for variable selection by tuning the penalization parameter through crossvalidation on AUC, thus avoiding overfitting. RESULTS: Fifty-three patients (28%) developed acute ≥ grade 3 dysphagia. The final model has an AUC of 0.71 and contains concurrent chemotherapy, D2 to the superior PC and the rs3213245 (XRCC1) polymorphism. The model's false negative rate and false positive rate in the optimal operation point on the ROC curve are 21% and 49%, respectively. CONCLUSIONS: This study demonstrated the utility of the EMLasso technique for model selection in predictive radiogenetics.
BACKGROUND AND PURPOSE: Design a model for prediction of acute dysphagia following intensity-modulated radiotherapy (IMRT) for head and neck cancer. Illustrate the use of the EMLasso technique for model selection. MATERIAL AND METHODS: Radiation-induced dysphagia was scored using CTCAE v.3.0 in 189 head and neck cancerpatients. Clinical data (gender, age, nicotine and alcohol use, diabetes, tumor location), treatment parameters (chemotherapy, surgery involving the primary tumor, lymph node dissection, overall treatment time), dosimetric parameters (doses delivered to pharyngeal constrictor (PC) muscles and esophagus) and 19 genetic polymorphisms were used in model building. The predicting model was achieved by EMLasso, i.e. an EM algorithm to account for missing values, applied to penalized logistic regression, which allows for variable selection by tuning the penalization parameter through crossvalidation on AUC, thus avoiding overfitting. RESULTS: Fifty-three patients (28%) developed acute ≥ grade 3 dysphagia. The final model has an AUC of 0.71 and contains concurrent chemotherapy, D2 to the superior PC and the rs3213245 (XRCC1) polymorphism. The model's false negative rate and false positive rate in the optimal operation point on the ROC curve are 21% and 49%, respectively. CONCLUSIONS: This study demonstrated the utility of the EMLasso technique for model selection in predictive radiogenetics.
Authors: Sarah L Kerns; Suman Kundu; Jung Hun Oh; Sandeep K Singhal; Michelle Janelsins; Lois B Travis; Joseph O Deasy; A Cecile J E Janssens; Harry Ostrer; Matthew Parliament; Nawaid Usmani; Barry S Rosenstein Journal: Semin Radiat Oncol Date: 2015-05-15 Impact factor: 5.934
Authors: D Alterio; M A Gerardi; L Cella; R Spoto; V Zurlo; A Sabbatini; C Fodor; V D'Avino; M Conson; F Valoriani; D Ciardo; R Pacelli; A Ferrari; P Maisonneuve; L Preda; R Bruschini; M Cossu Rocca; E Rondi; S Colangione; G Palma; S Dicuonzo; R Orecchia; G Sanguineti; B A Jereczek-Fossa Journal: Strahlenther Onkol Date: 2017-09-07 Impact factor: 3.621