In this study, a clinical dataset of patients with psychosis who have undergone Metacognitive Training (MCT) is analyzed using unsupervised machine learning (ML) methods to discover patient groups that are medically significant. The Positive and Negative Syndrome Scale (PANSS) is used in this study to assess the patient’s health state before and after MCT treatment as well as to measure treatment effectiveness. Four well-separated patient groups are found based on an exploratory data analysis with K-means clustering and outlier detection using Local Outlier Factor (LOF) algorithm. The statistical analysis of the improvement in positive, negative, and general symptoms for these patient groups provides meaningful insights about the patient’s profiles with regard to the effectiveness of the therapy. The results of the study confirm that the retrospective patient database under study comprises several differently-behaving groups of patients with a substantial variation in their data distribution, which needs to be addressed in the development of the predictive model for MCT effectiveness in future research.