Psychotic disorders pose a significant public health concern, and early detection of relapse is a crucial aspect of their management. In this study, we investigate the effectiveness of unsupervised learning-based anomaly detection approaches for relapse detection in psychotic disorders, using data from wearable sensors as a proposed solution for the ICASSP Signal Processing Grand Challenge e-Prevention track 2 Relapse Detection. We explore both a global scheme, where a single model is trained on data from all patients, and a personalized scheme, where a unique model is trained for each patient. To evaluate the performance of different anomaly detection models, we extract features related to sleep, heart rate, accelerometer and gyroscope data, physical activity, and percentage of time the sensor is worn. Additionally, we examine the influence of the time window size used for feature extraction. Our findings suggest that the autoencoder is the best-performing model in the global scheme while personalized models show significant variability in performance across patients and on average outperform the general model. These results highlight the potential of unsupervised learning approaches for relapse detection in psychotic disorders and underscore the importance of personalized approaches for optimal assessment of these conditions