#hope-s # Article Routine Clustering of Mobile Sensor Data Facilitates Psychotic Relapse Prediction in Schizophrenia Patients Shared by ![[Routine Clustering of Mobile Sensor Data Facilitates Psychotic Relapse Prediction in Schizophrenia Patients.pdf]]SOON on 13 Nov 2021 (Whatsapp Group Chat) #hope-s Joanne Zhou, Bishal Lamichhane, Dror Ben-Zeev, Andrew Campbell, Akane Sano AIM: Mobile sensing data to predict oncoming relapse of #schizophrenia patients to delivery timely #intervention Sample size: 63 patients. Monitor for 1 year. Clustering models: Gaussian Mixture Model, and Partition Around Medoids. Introduction on : [[What is Schizophrenia?]] **Mobile sensing can do [[Anomaly detection]] to predict relapse, that will help with relapse prevention and treatment. ** - Mobile sensing data has been used to model behaviors and predict psychotic relapses of schizophrenia patients in some of the earlier works. If an oncoming relapse could be detected with high accuracy then timely medical interventions could be provided to mitigate the associated risks. Researchers have found anomalies in daily behavior assessed from mobile sensing before relapses and developed relapse prediction models with promising accuracy [14–16]. In a pilot study, the Beiwe app collected mobile sensing data from 15 schizophrenia patients for 3 months during which 5 patients experienced relapses [14]. The researchers found that the rate of anomalies in mobility and social behavior increased significantly closer to relapses. In the CrossCheck project, a mobile sensing app was developed to collect self-reporting EMA (Ecological Momentary Assessment) and continuous passive sensing data from 75 outpatients with schizophrenia [17]. Based on this dataset, the authors in [15] compared different machine learning models for relapse prediction, with several feature extraction windows, and identified the best classifier and prediction settings for detecting an oncoming relapse. The best performance was obtained using an SVM (with RBF kernel) model and a feature extraction window of 30 days, leading to an F1 score of 0.27 on the relapse prediction task. Similarly, the authors in [18] used an anomaly detection framework based on an encoder-decoder reconstruction loss to predict psychotic relapse in schizophrenia. **One propose idea to monitor relapse is to monitor behavioral stability. ** ^131d2c - Concerning current mental health status, the extent to which an individual adheres to work, sleep, social, or mobility routine, i.e. a regular behavioral pattern, largely impacts their mental well-being and symptom severity of mental disorders [11,19,20]. Behavioral stability features have been proposed as relapse predictors in some of the previous studies. Features computed in our previous work measured behavioral stability by calculating the temporal evolution of daily templates of features derived from the mobile sensing data (daily templates are time-series obtained with representative feature values at regular time-intervals in a given day, e.g. time-series of hourly feature values) [16]. The authors in [21] also showed the effectiveness of using behavioral rhythm-based features to predict different symptom severity. Stability features such as deviation of daily templates were found to be significant predictors of schizophrenia symptoms such as being depressed. The authors in [22] also proposed a stability metric for behaviors with a fine temporal resolution by calculating the distance between two cumulative sum functions describing behaviors in a certain minute of the day. The computed Stability Index had similar predictive power as the state-of-the-art behavioral features (mean and standard deviation of each behavior) in [23], while being complementary. In all of these previous works utilizing behavioral stability to model relapse prediction, the stability measured was limited to the behaviors observed within a short feature extraction window (e.g. few weeks only). An individual’s routine behaviors were not fully represented due to the short time window considerations. A summary of behavioral patterns could rather be obtained when larger time windows are considered. ## Discussion ** The authors proposed clustering models to predict symptoms. I am not sure if it work yet. This is for the data scientists to understand. Too difficult for me at this point to understand. ** In this work, we proposed a methodology to compute clustering models on 24-hour daily behavior of schizophrenia outpatients, and showed that information extracted from the cluster model improved relapse prediction. New features were generated from the cluster models by measuring every observation’s deviation from the cluster centers representing typical behavioral patterns. Two different clustering models were investigated. The GMM model allows for cluster overlap and has a more extreme cluster dispersion. The PAM model with DTW distance creates partitional clusters that are more generalized towards new data but fails to identify dense clusters. The clustering-based features in addition to the baseline features helped to improve relapse prediction model performance. In future work, we will further investigate personalized clusters and relapse prediction models.