A Knowledge-Model for Predicting Radiation-Induced Esophagitis

Abstract

Purpose/Objective(s): Isotoxic improvements in treatment-outcome require the ability to predict radiationinduced complications. The main complications in (chemo-) radiotherapy are pneumonitis and esophagitis. The purpose of this study was to build a knowledge-model for predicting the probability of esophagitis incidence, based on clinical and treatment data. Materials/Methods: First, a literature review was done, and all relevant predictors with consistent results for esophagitis were identified. Based on these, a literature-based model was constructed. Second, data from 505 inoperable lung-cancer patients, treated with various schedules of (chemo-) radiotherapy were used to learn two different data-driven models for esophagitis, based on multivariate logistic regression and the relevance vector machine, respectively. While logistic regression used all the available predictors, the relevance vector machine was purely data-driven as it automatically selected the relevant predictors. Third, the literature-based and data-driven models were combined by an ensemble method (weighted mean) as to further improve prediction accuracy. Five-fold cross validation was used to obtain more reliable results. Performance of the models was expressed as the AUC (Area Under the Curve) of the Receiver Operating Characteristic (ROC) curve. The maximum value of the AUC is 1.0; indicating a perfect prediction model. A value of 0.5 indicates that patients are correctly classified in 50% of the cases, e.g. as good as chance. Results: The literature-based model for esophagitis consisted of concurrent vs. sequential vs. no chemotherapy, one vs. two fractions per day, and mean esophagus dose. The AUC was 0.78 on the data of the 505 patients. The logistic regression model consisted of gender, age, maximum esophagus dose, volume of esophagus, overall treatment time, total treatment dose, and biologically equivalent dose, additional to the above ones. The resulting cross-validated AUC was 0.78. The relevance vector machine automatically selected a set of predictors, including age, one or two fractions per day, mean esophagus dose, and biologically equivalent dose, and yielded a cross-validated AUC of 0.77. Combining the literature-based model (with weight 0.4) with the two data-driven models (with weights 0.3 for each), yielded a slightly improved AUC of 0.79. Conclusions: Combining demographic, imaging and treatment information yielded a knowledge model with accurate prediction of treatment-induced esophagitis. It appears feasible to build accurate models based solely on information from literature. Moreover, the literature model exclusively uses information known prior to treatment-start. Note that these predictions may be refined by using predictors measured during treatment, like the development of neutropenia in case of concurrent chemo-radiotherapy (De Ruysscher et al., Annals in Oncology, 2007). Our results suggest that this is a promising step toward isotoxic improvement of treatment-outcome of inoperable lung-cancer patients.