No abstract

Bipolar Disorder (BD) is a chronic mental illness characterized by changing episodes from euthymia (healthy state) through depression and mania to the mixed states. In this context, data collected through the interaction of patients with smartphones enable the creation of predictive models to support the early prediction of a starting episode. Previous research on predicting a new BD episode use mostly supervised learning methods that require labeled data and hence force a filtering of the available data to retain only those data that have valid labels (from the psychiatric assessment). To avoid limitations of supervised learning, in this paper we investigate the use of a semi-supervised learning approach that combines both labeled and unlabeled data to derive a model for BD episode prediction. Specifically we apply the DISSFCM (Dynamic Incremental Semi-Supervised Fuzzy C-Means) algorithm which offers the possibility to process in an incremental fashion the data stream of the voice signal captured by the smartphone, thus exploiting the evolving time structure of data which is ignored by static learning methods. DISS-FCM processes data in form of chunks and creates a dynamic collection of clusters thanks to a splitting mechanism that generates new clusters to better capture the hidden geometrical structure of data. This gives DISSFCM the ability to detect changes in data and dynamically adapt the model to them, thus improving the prediction accuracy. Preliminary results on real-world data collected at the Department of Affective Disorders, Institute of Psychiatry and Neurology in Warsaw (Poland) show that DISSFCM is able to predict some of healthy episodes (euthymia) and disease episodes even when only 25% of labeled data are available. Moreover DISSFM performs better than its previous version without split (ISSFCM) and it also overcomes the batch algorithm (SSFCM) that uses the whole dataset to create the model.