In this work, we have presented the sentiment analysis of messages posted on medical forums. We stated the sentiment analysis as a multi-class classification problem in which posts were classified into encouragement, gratitude, confusion, facts, facts + encouragement and uncertain categories. We applied the reader-centered manual annotation and achieved a strong agreement between the annotators: Fleiss Kappa = 0.73. We presented an ad-hoc method of the lexicon creation which is comparatively easy to implement. We have shown that the lexicon, which we call HealthAffect, provided the best accuracy in machine learning experiments. . We used two algorithms, NB and KNN, to solve a multi-class sentiment classification problem. The probability-based NB demonstrated a better performance than KNN.