*2.4. Discussion*

Related works aim to make association, detection, classification, and prediction about mental health. From the analysis of these studies, we identified some open issues about applying passive detection of social situations to support mental health professionals. First, it is necessary to develop solutions able to identify sociability patterns that represent information about social routine of individuals. Second, analyses of such solutions should consider contextual information to identify social situations. Finally, there is a need for solutions that can detect social behavior changes to allow specialized professionals to investigate whether there is a relationship between the identified change and the patient's mental state.

Although the related works aim to identify social behaviors to support mental health monitoring, these studies differ from our solution, so it is a challenge to use objective comparison metrics. For example, some studies develop machine learning models to classify and predict mental states, while our solution aims to extract sociability patterns and detect behavior changes. The works [22–24] also propose solutions capable of detecting sociability patterns, but they differ from our solution. These works design sociability patterns to quantify the duration and frequency of social interactions, while our solution recognizes periods of the day representing individuals' social routine. Besides, these works do not recognize sociability patterns based on contextual data (e.g., weekends, holidays, and rainy days) and do not perform incremental learning, then requiring to execute batch processing.

In comparison to related works, our research has the following contributions. First, this study does not focus on the diagnosis of a specific state or mental disorder but works on identifying situations of interest (i.e., the sociability routine) for mental health professionals. Second, the proposed solution recognizes context-sensitive sociability patterns, so enabling to distinguish normal behavioral variation from behaviors that are considered anomalies. Third, besides identifying the periods of the day when the individual generally socializes, the proposed solution can recognize unusual social routines and significant changes in the patient's social habit. Finally, the proposed solution uses a data stream mining approach to learn from social observations continuously.
