Exploring the Design of Conversational Agents for Mental Health Self-Monitoring

Abstract

An enduring aspiration is to use conversational agents (CAs) to provide mental health support (MHS) in order to tackle the persistent mental health treatment gap. CAs are computer interfaces that interact with people in a manner resembling human conversation. They are an alluring means of providing MHS as they can be designed to provide human like support in increasingly accessible, engaging, and personalised ways. The aspiration of using CAs to provide MHS is starting to be realised with CAs available today as smartphone applications. We need to understand how to design effective MHS CA systems so as to realise the potential to provide people with much needed support. If we do not learn how to design these systems effectively we risk missing out on the potential benefits and, worse, creating systems that cause harm. The design of CAs for MHS is a complex task with both CAs and MHS being innately complex. CAs are complex because of their human-likeness. Our understanding of how people interact with and desire to interact with these human-like computer interfaces is nascent. Mental health support is complex because people have diverse perspectives on the nature of mental health and how people should be supported with it. To navigate the complex design of CAs for MHS we need to develop understanding of the ways they could be designed to support people with their mental health. To develop understanding of how CAs could be designed to provide MHS, this thesis examines the design of CAs for the MHS activity of mental health self-monitoring (MHSM). MHSM is used to support people experiencing a range of mental health conditions including depression, anxiety, and bipolar disorder. It is however an innately challenging activity that many people struggle to do. CAs appear a promising technology to support people with MHSM and its challenging nature. If CAs can be designed to support people to do MHSM more effectively, they could benefit the mental health of many people. This thesis presents a series of three stands of work. In an effort to consider advanced utilisations of CAs for MHSM we first propose a plausible vision of future systems which we term pervasive therapy. This vision describes the general design of CA systems for the delivery of ecological momentary interventions (in the moment mental health support) as a component of blended care support (support combining human and digital support). We use this vision as a basis for discussion of three key design challenges: 1) intervention and conversation planning, 2) CA interaction design, and 3) clinician involvement. Furthermore we use this vision to discuss the ethical concerns and risks of pervasive therapy systems. Next this thesis examines the challenge of CA interaction design for MHSM. In order to conceptualise and explore the design of these interactions we draw on pertinent theories of human-human interaction, namely joint action theory and politeness theory. Building upon this we conducted a mixed-methods interview-based study (n=39) to investigate: 1) how the use of politeness by CA impacts the user experience, and 2) people's perspectives on the use of CAs for MHSM more broadly. This study demonstrates that the use of politeness by CAs for MHSM can both positively and negatively impact the user experience. Our analysis of participants' perspectives makes apparent conflicts between the perceived opportunities and risks of CAs for MHSM. This study raises questions about the ways in which it is appropriate and desirable for CAs to support people with their mental health. This thesis goes on to investigate the ways CAs could be designed to support people with MHSM. We draw on human-human social support theory in order to conceptualise and explore the design of CA support for MHSM. We conducted a qualitative interview study (N=24) building on this theory-based work to investigate people's perspectives on the ways CAs could be designed to support people with MHSM. Our interview design asked participants to consider the use of MHSM CAs by both themselves and a person close to them (e.g., partner, sibling) which enabled us to develop a nuanced understanding of their perspectives. We argue that participants reject CA support that they view as replacing human support, but are receptive to CA support that they view as augmenting or facilitating human support. To conclude, this thesis reflects on what has been learnt about the design of CAs for MHSM. This thesis ultimately argues that to design CA for MHSM responsibly we should move from dyadic views of individual users and CAs to a community view of users and CAs as part of larger support networks.