Understanding the potential and pitfalls of digital phenotypes to measure population mental health and wellbeing

Abstract

BackgroundDigital phenotypes such as social media data are increasingly being used to infer supposed trends in population mental health and wellbeing. These methods are attractive for their potential scale and timeliness. However, many methodological aspects need investigation to establish if these approaches could be applied successfully. We aim to explore these challenges by linking digital phenotypes to high-quality epidemiological data in a birth cohort study. MethodsWe retrospectively linked social media data from Twitter for 661 consenting participants in the Avon Longitudinal Study of Parents and Children, a multigenerational population cohort study. Participants volunteered their Twitter usernames, after a series of focus groups with 14 cohort participants about the acceptability of social media data linkage. Twitter data were used to test the effectiveness of inferring mental health and wellbeing from a range of sentiment analysis algorithms in combination with behavioural and temporal features. These methods were evaluated with participant-level survey data collected during the COVID-19 pandemic (April 9, 2020, to July 3, 2020), including with the Warwick Edinburgh Mental Wellbeing Scale, the General Anxiety Disorder-7 scale, and the Mood and Feelings Questionnaire. Ethical approval for the study was obtained from the Avon Longitudinal Study of Parents and Children Ethics and Law Committee, and informed consent was obtained from participants. FindingsTwitter provides a specific Application Programming Interface for academic research, granting free access to the full history archive and thus enabling the collection of data from cohorts. The focus groups determined that data linkage with Twitter and subsequent sharing is acceptable when data are anonymised. However, publicly available data such as Twitter data can be easily identifiable, and we are therefore exploring methods that would preserve privacy more effectively for data sharing. Our initial results show that temporal features have a key influence on the sensitivity and specificity of a model for the inference of wellbeing, anxiety, and depression when used in combination with textual and behavioural features. This influence varies depending on the specific outcome measure. Ongoing analyses will establish whether extrapolating successful models to alternative population samples is possible. InterpretationEarly results suggest that different mental health and wellbeing outcomes require individual models with uniquely tuned features. Cohort studies are a rich source of training data, but attention must be given to adequate anonymisation practices. FundingThis work was supported by the Health Foundation, the Alan Turing Institute, a Philip Leverhulme Prize to CMAH, and the Medical Research Council.