Social contagion? : using data science to characterise the distribution and dispersion of health behaviours in adolescence

Mental illness is the largest cause of disability in the UK with far-reaching consequences, spanning education, work, and health, across the life course. Analytical approaches for monitoring public mental health and evaluating the impact of policy and interventions - at scale - are urgently needed. In this work I will use data science methods to characterise health behaviours and their spread through social groups, with a particular focus on adolescent self-harm behaviours (injuries related to drug/alcohol-use or violence, and intentional self-injury) in schools. The work will also explore how linked health and education data can be used as a platform for randomised trials of health interventions in schools and hospitals.

In England, the average secondary school classroom includes three children who will ever self-harm, with at least one child who is admitted to hospital with self-inflicted or violent injuries aged 10-19 years. Self-harm in adolescence is predictive of premature death (particularly with suicide and drug/alcohol-use causes) and future A&E attendances and hospital admissions. The causes of self-harm are varied and highly complex, reflecting biological, social and personality factors in tandem with environmental triggers, which could be appropriate targets for intervention (e.g. exposure to others' self-injurious behaviour). We currently lack objective measures of the scale and clustering of self-harm behaviours in schools, because studies in this field have often used cross-sectional or panel (repeated survey) designs that do not capture the precise timing of events, or are not mapped to schools. Identification of predictors of peer-group and individual self-harm behaviours will inform targeted prevention strategies in schools and hospitals. The proposed work applies data science methods to a range of de-identified electronic health records and large-scale education datasets to characterise different self-harm presentations, and investigates the timing and sequence of these health behaviours within school peer-groups.

Health trajectories (e.g. future risk of death) will be characterised for different presentations of self-harm. The first step in this process is developing systematic approaches for identifying self-harm presentations ("clinical phenotypes") in a range of healthcare settings, including general practice, A&E and hospitals. Next, the influence of environmental influences (describing the "exposome") will be investigated as risk factors for clinical phenotypes of self-harm within social groups. This phase of work draws on linked health and education data for England and Scotland to create detailed information on the timing and sequence of health behaviours in schools peer-groups, to investigate evidence of social contagion. Data science methods (e.g. natural language processing) will be applied to extract information from free text in medical records, and a combination of machine learning, statistical and epidemiological methods will be used to develop algorithms for detecting the most clinically important presentations of self-harm in schools-peer groups.

The overarching aim of this work is to appraise and develop approaches for examining the distribution and dispersion of health behaviours captured in electronic health records. Self-harm behaviours associated with high-volume healthcare usage (intentional self-injury, substance misuse and violence) in secondary schools will be used as an exemplar. Three phases of work are planned. Phase 1 focuses on developing clinical phenotypes using data science methods to decipher the complexity of electronic health records. Phase 2 will appraise and apply methods for characterising the distribution and dispersion of clinical phenotypes in relation to the wider exposome, using linked health and education data. Phase 3 explores the potential for linked data to be used as a platform for enhancing the efficiency of randomised controlled trials of individually or cluster-randomised (at school or hospital-level) interventions. Data science methods will support the proposed work, in particular, natural language processing will support free text mining (e.g. free text recording of reason for A&E attendance) to enhance the development of clinical phenotypes for self-harm. Machine learning methods (e.g. random forest) will be explored as a means of classifying clinical subtypes and for investigating new approaches to characterising the interplay between the clinical phenome and exposome. Linked health and education data will be investigated as a platform for randomised controlled trials, and for nesting data capture with passive or sensing technology. Collectively the proposed work will establish tools for early phase research using linked health and education data, and provide new knowledge and insight into the dispersion of self-harm health behaviours in schools.