Machine learning methods are now widely used to study fMRI data in schizophrenia. One commonly studied brain network is the Default Mode Network (DMN), but its usefulness in tasks like classification is still not fully known. In this work, we tested if brain activity patterns from DMN regions, during an auditory oddball task, can help separate schizophrenia patients from healthy controls. We used data from the FBIRN Phase 2 3T dataset (25 schizophrenia and 25 controls). From the DMN areas defined by the Yeo atlas, we took voxel-wise contrast values (deviant > standard), and after selecting features and applying PCA, we trained logistic regression models. Models using only DMN features gave around 50% accuracy, close to chance. But models using whole-brain data performed better, with accuracy up to 80% and average cross-validated accuracy of 67.5%. Interpretation using SHAP showed that most useful voxels were not in the DMN but in visual and cerebellar areas. This suggests that DMN alone may not be enough for schizophrenia classification in task-based fMRI. Future work should focus on combining multiple brain networks.