Embodied Conversational Agents in Clinical Psychology

Following an internet-based therapy can be an effective way to deal with a depression. While some of these therapies can be followed independently, most of them still include some form of support by a therapist or coach. Interventions that do not require support are potentially more accessible and scalable, but thus far supported interventions appear to be a bit more effective. In this project we aim to bridge the gap between guided and unguided interventions by automating human support, or part of it, with embodied conversational agents.

You might know embodied conversational agents (ECAs) under a different name, for example, avatar or virtual character. ECA is an academic term for the concept for which the website www.chatbots.org/ has already identified 161 synonyms, summarized as “humanlike conversational AI entities”. Since this does not exactly make things more clear, a definition is warranted. We are talking about computer programs that (1) have an embodiment, for example on a computer screen, (2) can communicate with users in a human-like manner, and (3) apply artificial intelligence (AI) to show intelligent behavior. Both the balance between these three elements, as well as the complexity with which they can be implemented can differ tremendously. Consider, for example, the difference between a chatbot, represented by a picture, that answers questions in an online store, and a very realistic video gaming character with whom a user can barely interact.

From ELIZA to Ellie
The classical example in psychology is chatbot ELIZA, developed back in 1966 by Joseph Weizenbaum, which simulates a Rogerian therapist. It is still possible to talk to ELIZA if you go to www.masswerk.at/elizabot/. One of the things ELIZA has a lot of trouble with, just like computers in the present, is interpreting semantics. For this reason we are probably still far removed from computers taking over the clinical interviews, and psychologists becoming obsolete. Nevertheless, the field has naturally progressed a lot over the past 50 years, and especially the following Youtube video of virtual counsellor ‘Ellie’ appeals to the imagination: https://www.youtube.com/watch?v=ejczMs6b1Q4. Can we use this kind of technology in routine clinical practice?

Literature review
In our first study we disclosed the research field with a review of the literature. We identified a total of 49 studies that used an ECA in an intervention for people with common mental health disorders. One of our primary interest was the level of evidence that supported the interventions that we found. Had they already been shown to be effective and safe for application in routine practice, was there perhaps some solution with which we could immediately get started, or would we have to follow our own path?

Relatively new
What we found was that most of the interventions were still in the development and piloting phases. The studies mostly dealt with trying out new ideas, and testing their feasibility. Few studies answered important questions that arise when we want to start using new technologies in routine practice: Do symptoms improve? Is it safe to use? What are the long-term effects? Are ECA interventions more effective than the interventions we already have? Of course, this is hardly surprising if we consider that we are dealing with a relatively new field of research; it was only after 2009 that we started seeing an increase in the number of studies that was published.

Simple but effective
Additionally, developing ECA systems such as Ellie is not a trivial task. First of all experts from different disciplines are necessary, for example, computer scientists and psychologists. Furthermore, it can take years to develop the different components that make up an ECA, such as dialog engines, or automated non-verbal behavior. Because the systems can grow very complex, it becomes harder to convince both ethical boards and practitioners to start testing them with real patients. In order to contribute to the evidence base in the limited time-span of a PhD trajectory, we decided to go with a more ‘low-tech’ approach; virtual characters that still match all three criteria mentioned before, but are a lot easier to develop than, for example, Ellie. Furthermore, we chose to limit ourselves to a task and outcome measure that are important in many of our interventions: Ecological Momentary Assessment (EMA), and adherence.

Ecological Momentary Assessment and Adherence
EMA refers to repeated measurements of people’s behavior or experiences, in their own environment, and in the current moment. A good example is the paper diary in which people with depression register their mood over the course of a week, for example, to see whether there might certain recurring events that trigger bad moods. These days paper diaries are being replaced by smartphone applications that are able to send reminders at specific times. In EMA, adherence, or compliance, is very concrete: people either respond or do not respond to the automated requests.

Motivation
This is what we looked at in our second study, in which we investigated whether we could increase adherence with a simple form of visual feedback, and whether we should take into account individual differences in people’s motivation. In a three-week smartphone study, participants reported their mood three times a day in an EMA app. Half of the participants were thanked for their response by an avatar that mirrored their self-reported mood state. Our goal was to see whether lowly motivated participants would become increasingly motivated by a system in which they could recognize themselves. Surprisingly, not the lowly motivated, but the highly motivated participants were the ones to benefit from the feedback in this sense. Thus, an important lesson we learned from this study is that merely giving feedback does not necessarily lead to the desired results, and that it could even work out the other way around.

The next step in this project is the development of a new and generic EMA application. Within this application there will be more room for experimenting with automated feedback, for example through personalization. Building on the paradigm of our second study, our goal is to find out whether inclusion of automated feedback can increase adherence.