Using smartphone-based personal sensing to understand and predict risk of psychotic relapse at the individual level.

Despite the relatively high efficacy of antipsychotic medication in the acute stage of psychosis, it is estimated that around 84% of individuals will go on to experience at least one further episode within the first 36 months (1). Psychotic relapse causes considerable distress not only to the patients but also their families. Furthermore, psychotic relapse often results in hospitalisation and is therefore associated with significant health care costs. Additionally, each relapse appears to have a negative effect on the underlying neurobiology of the disorder, leading to a worsening of long-term outcomes (2). Relapse prevention is therefore one of the most critical targets in the treatment of psychotic disorders (3).The ability to predict psychotic relapse would provide the opportunity for clinicians to tailor interventions and to apply them effectively and efficiently to individuals who are at highest risk (4). A risk score which quantifies the probability of psychotic relapse and hospital readmission on an individual level could be used to identify individuals at most need of an intervention (5).However, we currently have a very limited understanding of factors that predict psychotic relapse and are therefore unable to stratify patients based on their levels of risk.Using machine learning and smartphone-based data to predictpsychotic relapseRecent advances in machine learning (ML) methods have the potential to revolutionise psychosis prediction (6). ML is particularly useful when aiming to predict relapse on an individual rather than group level, leading to greater potential of translation into clinical practice where clinicians need to make decisions about individual patients. In addition, ML models are typically multivariate andcan therefore capture the hidden relationships between different predictive variables.There are several widely-used ML techniques, divided into 3 broad categories: supervised, unsupervised and semi-supervised (7). The most suitable technique depends on the task to be performed and the characteristics of the data, such as sample size and dimensionality.Smartphones are particularly suitable as a tool for data collection given their increasing ubiquity and flexibility. In addition,they place minimal burden on participant and can be particularly useful when participants are less engaged and harder to reach which is when they are at greater need for support and monitoring.Work leading to the present investigationThe Urban Mind app (https://www.urbanmind.info/) has been developed over the past 5 years and has been successfully piloted in the general population (8)as well as in clinical populations (first-episode psychosis and ultra-high risk for psychosis). It collects active data in the form of Ecological Momentary Assessments (EMA) and passive data such as GPS location and number of steps per day. EMA methodology involves sampling participants experiences in real time and in real-world contexts (9). Participants are asked about how they feel in the moment, minimising the risk of recall bias. In the Social Mind study, an adapted version of the app will be used, with specific focus on social environment and social stress.