When a human makes a decision, an observer may want to understand the reasons and motivations behind the decision. This understanding is important when IVAs are involved in contextual decision-making or coaching practices. To address this challenge, we propose that an agent’s understanding of its user should include knowledge of the user’s underlying values. Humans prioritise different values – sometimes contradictory – in a manner that depends on the context. We present a method where the agent and user build the required context-sensitive value model together. We use Schwartz’s value theory, which places individuals’ values into ten categories. In a between-subject experiment, with three sessions on different days, we elicit user values by presenting them with moral dilemmas in different contexts on the first day, refine the model by asking users to argue about contradictions on the second day, and let them reflect on the model that they have built together with the system on the third day. We find that users exposed to a value-aware condition are more likely to agree with the robot’s representations of their values post-reflection than those in a baseline. Participants also prioritise different values depending on the context, agreeing with previous findings.@en

E-mental health for depression is increasingly used in clinical practice, but patient adherence suffers as therapist involvement decreases. One reason may be the low responsiveness of existing programs: especially autonomous systems are lacking in their input interpretation and feedback-giving capabilities. Here, we explore (a) to what extent a more socially intelligent and, therefore, technologically advanced solution, namely a conversational agent, is a feasible means of collecting thought record data in dialog, (b) what people write about in their thought records, (c) whether providing content-based feedback increases motivation for thought recording, a core technique of cognitive therapy that helps patients gain an understanding of how their thoughts cause their feelings. Using the crowd-sourcing platform Prolific, 308 participants with subclinical depression symptoms were recruited and split into three conditions of varying feedback richness using the minimization method of randomization. They completed two thought recording sessions with the conversational agent: one practice session with scenarios and one open session using situations from their own lives. All participants were able to complete thought records with the agent such that the thoughts could be interpreted by the machine learning algorithm, rendering the completion of thought records with the agent feasible. Participants chose interpersonal situations nearly three times as often as achievement-related situations in the open chat session. The three most common underlying schemas were the Attachment, Competence, and Global Self-evaluation schemas. No support was found for a motivational effect of providing richer feedback. In addition to our findings, we publish the dataset of thought records for interested researchers and developers.
@en

This thesis investigates how artificial intelligence can support e-mental health for depression, i.e. the delivery of treatment and prevention interventions for depression using technology. E-mental health for depression is a promising means for bridging the treatment gap since it addresses many of the barriers that prevent people in need of help from seeking or obtaining it. Additionally, many systems have been found to be effective in controlled trials. However, as human support for e-health interventions decreases so do their effectiveness and users’ adherence. While one possible explanation is that human support is a necessary ingredient of a successful intervention, another is that the technology is not satisfying the needs of users to the best of its abilities. This finding inspired us to take a closer look at the technological implementation of the functionality of these systems. To this end, we developed a set of scales that assess the technological sophistication of the functional components of systems, the e-mental health degree of technological sophistication (eHDTS) scales. In a systematic literature review of the field, we then divided all systems developed between 2000 and 2017 for the prevention or treatment of depression reported in the scientific literature into their functional components and rated those components with the eHDTS scales. We found that most systems that had been developed until 2017 were low-tech implementations, consisting mostly of psychoeducation and having a one-way information stream from system to user. This clearly contrasts with face-to-face therapy in which the therapist closely attends to the patient and provides his or her knowledge and insight strategically to signal understanding and empathy, foster self-reflection, teach, or obtain more information. Based on this consideration, we set out to develop a conversational agent capable of signaling to the user that it had processed the content of what it had been told when completing a thought record together with a user in dialog with the hypothesis that this would be able to motivate the user to complete more thought records and feel more engaged. Thought recording is a core technique of cognitive therapy in which patients are asked to systematically monitor their thinking in situations that caused a maladaptive response. Cognitive theory posits that the negative, cognitive appraisals that are responsible for the low mood experienced in patients with depression stem from maladaptive schemas, i.e., beliefs that we hold as truths about the world, ourselves, and the future. To get the conversational agent to “understand” the thoughts provided by the user from this cognitive theory perspective, we collected a corpus of thought records from Amazon Mechanical Turk workers, manually coded the thoughts with respect to the underlying schema, and trained various machine learning models to do the same labeling. A set of deep neural networks outperformed the other algorithms and was then deployed in the conversational agent. We used a between-subjects design to expose 308 participants recruited from Prolific to the conversational agent. The three conditions differed with respect to the feedback-giving capabilities of the conversational agent in response to a thought record: low feedback richness entailed an acknowledgment of the completion of the thought record (thanking the user), medium feedback richness entailed the acknowledgment plus feedback on the process (how many steps the user did in relation to his or her previous thought records), and rich feedback richness entailed medium feedback richness combined with feedback on the content (an interpretation of the thought record with respect to the underlying schema). While all users were able to complete the thought records with the conversational agent, we did not find supportive evidence that the agent’s feedback strategy could increase users’ motivation to complete more thought records or their self-reported engagement in self-reflection. Future research may investigate why we observed these null results by studying whether the feedback is processed correctly, whether a population with depression that is motivated by a wish to get healthy might behave or experience the system differently from our sample that was recruited online and did not meet diagnostic criteria for depression, or whether more advanced social and interaction capabilities need to accompany the complex feedback for it to be believable.@en

The cognitive approach to psychotherapy aims to change patients’ maladaptive schemas, that is, overly negative views on themselves, the world, or the future. To obtain awareness of these views, they record their thought processes in situations that caused pathogenic emotional responses. The schemas underlying such thought records have, thus far, been largely manually identified. Using recent advances in natural language processing, we take this one step further by automatically extracting schemas from thought records. To this end, we asked 320 healthy participants on Amazon Mechanical Turk to each complete five thought records consisting of several utterances reflecting cognitive processes. Agreement between two raters on manually scoring the utterances with respect to how much they reflect each schema was substantial (Cohen’s κ = 0.79). Natural language processing software pretrained on all English Wikipedia articles from 2014 (GLoVE embeddings) was used to represent words and utterances, which were then mapped to schemas using k-nearest neighbors algorithms, support vector machines, and recurrent neural networks. For the more frequently occurring schemas, all algorithms were able to leverage linguistic patterns. For example, the scores assigned to the Competence schema by the algorithms correlated with the manually assigned scores with Spearman correlations ranging between 0.64 and 0.76. For six of the nine schemas, a set of recurrent neural networks trained separately for each of the schemas outperformed the other algorithms. We present our results here as a benchmark solution, since we conducted this research to explore the possibility of automatically processing qualitative mental health data and did not aim to achieve optimal performance with any of the explored models. The dataset of 1600 thought records comprising 5747 utterances is published together with this article for researchers and machine learning enthusiasts to improve upon our outcomes. Based on our promising results, we see further opportunities for using free-text input and subsequent natural language processing in other common therapeutic tools, such as ecological momentary assessments, automated case conceptualizations, and, more generally, as an alternative to mental health scales.@en

BACKGROUND: Electronic mental (e-mental) health care for depression aims to overcome barriers to and limitations of face-to-face treatment. Owing to the high and growing demand for mental health care, a large number of such information and communication technology systems have been developed in recent years. Consequently, a diverse system landscape formed. OBJECTIVE: This literature review aims to give an overview of this landscape of e-mental health systems for the prevention and treatment of major depressive disorder, focusing on three main research questions: (1) What types of systems exist? (2) How technologically advanced are these systems? (3) How has the system landscape evolved between 2000 and 2017? METHODS: Publications eligible for inclusion described e-mental health software for the prevention or treatment of major depressive disorder. Additionally, the software had to have been evaluated with end users and developed since 2000. After screening, 270 records remained for inclusion. We constructed a taxonomy concerning software systems, their functions, how technologized these were in their realization, and how systems were evaluated, and then, we extracted this information from the included records. We define here as functions any component of the system that delivers either treatment or adherence support to the user. For this coding process, an elaborate classification hierarchy for functions was developed yielding a total of 133 systems with 2163 functions. The systems and their functions were analyzed quantitatively, with a focus on technological realization. RESULTS: There are various types of systems. However, most are delivered on the World Wide Web (76%), and most implement cognitive behavioral therapy techniques (85%). In terms of content, systems contain twice as many treatment functions as adherence support functions, on average. Furthermore, autonomous systems, those not including human guidance, are equally as technologized and have one-third less functions than guided ones. Therefore, lack of guidance is neither compensated with additional functions nor compensated by technologizing functions to a greater degree. Although several high-tech solutions could be found, the average system falls between a purely informational system and one that allows for data entry but without automatically processing these data. Moreover, no clear increase in the technological capabilities of systems showed in the field, between 2000 and 2017, despite a marked growth in system quantity. Finally, more sophisticated systems were evaluated less often in comparative trials than less sophisticated ones (OR 0.59). CONCLUSIONS: The findings indicate that when developers create systems, there is a greater focus on implementing therapeutic treatment than adherence support. Although the field is very active, as evidenced by the growing number of systems developed per year, the technological possibilities explored are limited. In addition to allowing developers to compare their system with others, we anticipate that this review will help researchers identify opportunities in the field.

@en

An important aspect of human emotion perception is the use of contextual information to understand others' feelings even in situations where their behavior is not very expressive or has an emotionally ambiguous meaning. For technology to successfully detect affect, it must mimic this human ability when analyzing audiovisual input. Databases upon which machine learning algorithms are trained should capture the context of social interactions as well as the behavior expressed in them. However, there is a lack of consensus about what constitutes relevant context in such databases. In this article, we make two contributions towards overcoming this challenge: (a) we identify two principal sources of context for emotion perceptions based on psychological theory, and (b) we provide an overview of how each of these has been considered in published databases covering social interactions. Our results show that a similar set of contextual features are present across the reviewed databases. Between all the different databases researchers seem to have taken into account a set of contextual features reflecting the sources of context seen in psychological theory. However, within individual databases, these features are not yet systematically varied. This is problematic because it prevents them from being used directly as resources for the modeling of context-sensitive affect detection. Based on our findings, we suggest improvements for the future development of affective databases.

@en

To date, meta-analyses of e-mental health systems for major depressive disorder (MDD) have largely overlooked the technological side of interventions. This warranted the creation of an open access database, EHealth4MDD, for the systematic study of the technological implementation in relation to intervention content, study design, and study outcomes. E-health systems were identified by conducting an exhaustive search on PubMed, Scopus, and Web of Science in 2017. The 5379 retrieved records yielded 267 systems. One coder extracted information from the records on 45 variables, organized into 14 tables in EHealth4MDD. A sample of each high-inference variable was double coded by a second coder to assess reliability. Percent agreement was satisfactory given that coders received no training and the number of possible categories was large. Furthermore, scales were developed to rate the degree of technological sophistication of system functions for each of five function types. Four of these scales demonstrated concurrent validity, as evidenced by the substantial to strong correlations observed when comparing the scales with the results of an unlabeled ordering task. For researchers in both computer science and clinical psychology, the database presents a useful tool to systematically study e-mental health interventions for depression.@en

A number of negotiation training systems have been developed to improve people’s performance in negotiation. They mainly focus on the skills development, and less on negotiation understanding and improving self-efficacy. We propose a virtual reality negotiation training system that exposes users to virtual cognitions during negotiation with virtual characters with the aim of improving people’s negotiation knowledge and self-efficacy. The virtual cognitions, delivered as a personalized voice-over, provide users with a stream of thoughts that reflects on the negotiation and people’s performance. To study the effectiveness of the system, a pilot study with eight participants was conducted. The results suggest that the system significantly enhanced people’s knowledge about negotiation and increased their self-efficacy.

@en

A key challenge in developing companion agents for children is keeping them interested after novelty effects wear off. Self Determination Theory posits that motivation is sustained if the human feels related to another human. According to Social Penetration Theory, relatedness can be established through the reciprocal disclosure of information about the self. Inspired by these social psychology theories, we developed a disclosure dialog module to study the self-disclosing behavior of children in response to that of a virtual agent. The module was integrated into a mobile application with avatar presence for diabetic children and subsequently used by 11 children in an exploratory field study over the course of approximately two weeks at home. The number of disclosures that children made to the avatar during the study indicated the relatedness they felt towards the agent at the end of the study. While all children showed a decline in their usage over time, more related children used the application more, and more consistently than less related children. Avatar disclosures of lower intimacy were reciprocated more than avatar disclosures of higher intimacy. Girls reciprocated disclosures more frequently. No relationship was found between the intimacy level of agent disclosures and child disclosures. Particularly the last finding contradicts prior child-peer interaction research and should therefore be further examined in confirmatory research.

@en

Reciprocal self-disclosure is an integral part of social bonding between humans that has received little attention in the field of human-agent interaction. To study how children react to self-disclosures of a virtual agent, we developed a disclosure intimacy rating scale that can be used to assess both the intimacy level of agent disclosures and that of child disclosures. To this end, 72 disclosures were derived from a biography created for the agent and rated by 10 university students for intimacy. A principal component analysis and subsequent k-means clustering of the rated statements resulted in four distinct levels of intimacy based on the risk of a negative appraisal and the impact of betrayal by the listener. This validated rating scale can be readily used with other agents or interfaces.@en