Purpose/Objective(s): Multiple studies based on 18F-flourodeoxyglucose positron-emission tomography (FDG-PET) have reported correlations between simple SUV-based measures and outcomes in oropharyngeal cancer (OC). We investigated the use of shape and intensity feature datamining (‘radiomics’) and machine learning techniques to develop predictive models for local failure (LF), regional failure (RF), or distant metastasis (DM) following definitive chemoradiation therapy (CRT). Materials/Methods: From December 2002 to March 2009, all stage III-IV OC patients treated at our institution with CRT with retrievable pretreatment and post-treatment PET scans were identified (N Z 174). Pretreatment PET scans were converted to Computational Environment for Radiotherapy Research (CERR) format and 24 representative shape and intensity features of FDG-avid disease regions were extracted for each scan. Using machine learning feature selection methods, multiparameter logistic regression models were built incorporating clinical factors and imaging features. Rates of LF, RF, and DM were analyzed using logistic regression and Cox proportional-hazards regression with competing-risks analysis. Confidence intervals are 95%. Models were characterized by the Spearman rank correlation coefficient (Rs) and area under the ROC curve (AUC). All model-building methods were tested for significant correlation by leave-one-out cross-validation (LOOCV), and for LF with an independent patient cohort from a collaborating institution. Results: Median follow-up time was 55 months (5-112 months). Five-year estimated LF, RF, and DM rates were 7.4% (CI 3.2-11.6%), 6.1% (CI 2.39.9%), and 19.6% (13.4-25.8%), respectively. Smoking history (69% were long time smokers) was not significant. LF correlated with image intensity skewness and metabolic tumor volume (MTV), model Rs Z 0.277 (p < .001), AUC Z 0.826. RF correlated with shape heterogeneity and MTV, model Rs Z 0.283 (p < .001), AUC Z 0.866. DM correlated with image intensity kurtosis and MTV, model Rs Z 0.334 (p < .001), AUC Z 0.744. All model building methods produced significant correlations on LOOCV, with model Rs Z 0.207, 0.248, and 0.238 (p Z .007, .001, and .002) and AUC Z 0.743, 0.820, and 0.674 for LF, RF, and DM; the model for LF retained significant correlation in the independent patient cohort, Rs Z .186 (p Z .046), AUC 0.625. Conclusions: In the largest study of its kind to date, predictive models constructed using FDG-PET intensity and shape features significantly correlated to LF, RF, and DM in the test cohort and on LOOCV, and with LF with an independent population. Such models could assist in patient selection for dose reduction or escalation, or to identify high-risk patients for additional adjuvant therapy, but further understanding of the biological basis of image features will be necessary. Author Disclosure: M.R. Folkert: None. J.H. Oh: None. J. Setton: None. A.P. Apte: None. W.L. Thorstad: None. H. Schoder: None. N.Y. Lee: None. J.O. Deasy: E. Research Grant; Varian Medical Systems.