Characterisation of mental health conditions in social media using Informed Deep Learning

Mental and substance use disorders are the leading cause of years lived with disability worldwide and in 2010 accounted for 7.4% of years of productive life lost due to disability1. Natural language processing of electronic health records (EHRs) is increasingly being used to study mental illness2 and risk behaviours in much closer detail than previously3. However, narrative notes are written by clinicians who record those positive findings and relevant negatives that guide their subsequent diagnosis and treatment plan for the patient4. Although EHRs allow clinicians to synthesise disparate facts making them interpretable by other clinicians, they do not "paint a full picture" of the patient experience of a mental health problem, particularly as patients may answer interview questions in a manner that they perceive will be viewed favourably by their clinician. Moreover, as patient records are only written based on meetings with their healthcare provider, critical changes in patient behaviour and wellbeing may not be recognised either immediately or at all due to a time delay in reporting, thus preventing certain real time interventions.