We propose and evaluate a systematic approach to detect and classify Patient/Problem, Intervention, Comparison and Outcome (PICO) from the medical literature. The training and test corpora were generated systematically and automatically from structured PubMed abstracts. 23,472 sentences by exact pattern match of head words of P-I-O categories. Afterward, the terms with top frequencies were used as the features of Naïve Bayesian classifier. This approach achieves F-measure values of 0.91 for Patient/Problem, 0.75 for Intervention and 0.88 for Outcome, comparable to previous studied based on mixed textural, paragraphical, and semantic features. In conclusion, we show that by stricter pattern matching criteria of training set, detection and classification of PICO elements can be reproducible with minimal expert intervention. The results of this work are higher than previous studies.