Using Responsible Artificial Intelligence (AI) to Predict Online Therapy Outcome and Engagement

Mental disorders contribute greatly to the global disease burden, but many people do not have access to mental health care. This treatment gap is partly due to structural (e.g., availability) and attitude-related (e.g. fear of stigma) barriers in health care seeking. Digital therapeutics (DTx) in the form of digital mental health interventions or digital psychotherapy may be the solution to this problem. The integration of Information and Communication Technology (ICT) and mental health care has the potential to increase the efficiency of care delivery and enables personalisation of treatments. Artificial Intelligence (AI)-based analysis of large datasets from digital psychotherapy programs may allow developing and validating personalised prediction models. The prediction of individual engagement and the early identification of untoward engagement patterns may improve personalisation of DTx, which could help reduce nonadherence and improve treatment outcome. The personalised prediction of DTx outcomes and engagement patterns may be achieved by implementing AI-based approaches, such as Machine Learning prediction models. Personalised prediction models may lead to a better understanding of who profits most from what kind of DTx in a real-world setting. Taken together, personalisation of DTx treatment outcomes and engagement may i) improve decision making processes in patient-clinician dyads, ii) improve efficiency of digital psychotherapy, iii) reduce suffering of patients, and iv) reduce direct and indirect cost related to mental health care. There is a need to account for potential discrimination due to mental health in AI-based predictions models. Unbiased and non- discriminating AI is often referred to as responsible AI. Accounting for bias in AI-based prediction models based on a specific dataset is especially important in mental health care to prevent acceleration of health discrimination.

This study is to develop AI-based models for the personalised prediction of treatment engagement and treatment outcomes in patients engaging in digital psychotherapy. A large, real-world dataset of patients in a digital psychotherapy program will be used to train AI algorithms. Responsible AI algorithms will be developed by describing, accounting for, and mitigating bias due to severity of mental disturbances in AI-based models, in addition to considering bias due to other sensitive attributes, such as gender, ethnicity, and socio-demographic status. The aim of the proposed project is to estimate AI-based prediction models of treatment engagement and outcomes based on data from the Online Therapy Unit by Prof. Heather Hadjistavropoulos from the University of Regina, Canada. The Online Therapy Unit dataset contains a large amount of data on DTx from people with mental disorders (collected as part of research trials in the Online Therapy Unit from 2013 to 2021) and is derived from the publicly funded, internet-delivered, cognitive behaviour therapy (iCBT) program in Saskatchewan, Canada. In sum, the Online Therapy Unit dataset is highly suitable as a training and test dataset for AI-based prediction models, as it comprises a large number of participants, longitudinal data retrieved from the real world opposed to a clinical trial, and a rich set of predictive features.