The topic of people's health has always attracted the attention of public and private structures, the patients themselves and, therefore, researchers. Social networks provide an immense amount of data for analysis of health-related issues; however it is not always the case that researchers have enough data to build sophisticated models. In the paper, we artificially create this limitation to test performance and stability of different popular algorithms on small samples of texts. There are two specificities in this research apart from the size of a sample: (a) here, instead of usual 5-star classification, we use combined classes reflecting a more practical view on medicines and treatments; (b) we consider both original and noisy data. The experiments were carried out using data extracted from the popular forum AskaPatient. For tuning parameters, GridSearchCV technique was used. The results show that in dealing with small and noisy data samples, GMDH Shell is superior to other methods. The work has a practical orientation.