In this work, we studied the efficiency of principal component analysis for feature extraction and classification of prostate cancer patients suffering from rectal bleeding. We fully exploited the three-dimensional planned dose distribution by considering the voxels as observations. We compared different possibilities for selecting the most relevant features (sequential and combinatory). The receiving operator characteristics were used as performance criterion. The obtained results demonstrate the ability of the method to classify two groups of patients, namely rectal bleeding and non-rectal bleeding. They also suggest that local dose/toxicity relationships exist.