In a recent review published in npj Digital Medicine, researchers examined the current state of research on fully automated interventions mediated by conversational agents (CAs) for the emotional component of mental health among young people.
survey: Using automated conversational agents to improve the mental health of young people: a scoping review. Image Credit: SewCreamStudio/Shutterstock.com
Background
Mental health problems are a serious concern for young people, leading to psychosocial difficulties in adulthood.
The technology has emerged as an alternative to face-to-face approaches, with CAs being digital solutions that simulate human interaction using text, speech, gestures, facial expressions or sensory expressions.
However, fully automated CAs have limitations, such as relying primarily on elderly populations and not distinguishing between young and older populations. Most reviews focus on a subcategory of conversational agents based on performance level.
About the review
In the current review, researchers explored the potential of automated conversational agents to improve youth psychiatric well-being.
The researchers searched PubMed, Web of Science, PsychInfo, Scopus, the Association for Computing Machinery (ACM) Digital Library, and IEEE Xplore in March 2023.
These included primary research studies reporting on the development, usability/feasibility or evaluation of fully autonomous conversational agents to improve the mental health of persons aged ≤25 years. All studies belong to peer-reviewed English-language journals.
The team excluded secondary research, dissertations, conference proceedings, and commentaries describing or reporting on general characteristics of human–conversational agent interactions or intervention studies exclusively testing general characteristics of human–technology interactions using CA.
They also excluded studies on applications of CA to improve cognitive, social, physical, or educational health and those that emphasized the use of CA only for monitoring or evaluation purposes. In addition, they excluded studies using semi- or non-automated CAs targeting individuals over 25 years of age.
Two independent investigators reviewed the records, and a third investigator resolved disagreements. Data extracted included general, technological, interventional and peer-reviewed study characteristics.
General aspects include year of publication, country and authors, while technological aspects include type of conversational system, name, modality of communication, availability and type of performance.
Intervention characteristics assessed included target mental health outcome, scope, frequency, duration, theoretical framework, or stand-alone intervention).
Study characteristics include information about participants, study methodology and design, study phase, and main outcomes.
Results
Of the 9,905 records initially identified, 6,874 were subject to title screening and 152 were subject to full text screening. However, only 25 eligible records including 1707 individuals were analyzed.
A total of 21 agents were identified, most of which were disembodied chatbots, robots, and virtual representations, of which most studies used Paro, Nao, and Woebot.
The dialog system used by the certification bodies was mostly machine learning and natural language processing (n=12), with some using predefined dialog systems and interactions mapped and assembled for dynamic user input.
Most CAs targeted anxiety (n=12), followed by depression, mental well-being, general distress, and mood. Most records refer to conversational agent applications as interventions, focusing on preventive measures for the general public and individuals at risk.
Nineteen studies reported the duration of the interventions, most lasting two to four weeks (eight studies). Seventeen studies reported theoretical frameworks for the interventions, with cognitive behavioral theory (CBT) applying to most interventions, and 14 automated CA applications mentioning positive psychology as their framework.
Other theories include interpersonal theory, person-centered theory, metacognitive narrative imagery intervention, motivational interviewing, transtheoretical approach, dialectical behavioral theory, and emotion-focused theory.
The sample size ranged from eight to 234 participants, recruited primarily from educational, community and health settings, with a mean age of 17 years and 58% female.
Fifteen studies reported feasibility outcomes including engagement, retention/adherence, acceptability, user satisfaction, system usability, safety, and functionality.
Two studies reported safety issues, with >50% of subjects reporting at least one adverse effect despite high feasibility. Fifteen studies reported anxiety outcomes, with five reporting a significant positive difference compared to controls.
A randomized controlled trial found an improvement in anxiety related to medical procedures for participants undergoing more invasive procedures and with more frequent exposure to medical procedures.
Nine studies reported depression, with five showing a significant difference compared to controls in favor of automated CAs.
In uncontrolled studies, one showed minimal change in depression scores and two studies showed a significant improvement in psychological well-being but no significant effect on subjective happiness.
Conclusion
In conclusion, based on the review findings, automated CAs can improve mental health outcomes, particularly for anxiety and depression; however, further research could improve understanding of their effectiveness and potential limitations.
The field is rapidly expanding with advanced technical capabilities, especially in high-income countries.
Future reviews should include a safety study, address a wide range of clinical issues, include larger sample sizes, and conduct cost-effectiveness studies to inform accessibility in low- and middle-income countries.