Background and objective: Classifying people according to their health profile is crucial in order to propose appropriate treatment. However, the medical diagnosis is sometimes not available. This is for example the case in health insurance, making the proposal of custom prevention plans difficult. When this is the case, an unsupervised clustering method is needed. This article aims to compare three different methods by adapting some text mining methods to the field of health insurance. Also, a new clustering stability measure is proposed in order to compare the stability of the tested processes. Methods : Nonnegative Matrix Factorization, the word2vec method, and marginalized Stacked Denoising Autoencoders are used and compared in order to create a high-quality input for a clustering method. A self-organizing map is then used to obtain the final clustering. A real health insurance database is used in order to test the methods. Results: the marginalized Stacked Denoising Autoencoder outperforms the other methods both in stability and result quality with our data. Conclusions: The use of text mining methods offers several possibilities to understand the context of any medical act. On a medical database, the process could reveal unexpected correlation between treatment, and thus, pathology. Moreover, this kind of method could exploit the refund dates contained in the data, but the tested method using temporality, word2vec, still needs to be improved since the results, even if satisfying, are not as better as the one offered by other methods.