Artificial intelligence (AI) chatbots are capable of mimicking human interactions using spoken, written or verbal communication with the user. AI chatbots can provide important health-related information and services, ultimately leading to promising technology-facilitated interventions.
study: Artificial intelligence (AI)-based chatbots to promote behavioral health changes: a systematic review. Image credit: TippaPatt / Shutterstock.com
AI chatbots in healthcare
Current digital telehealth and therapeutic interventions are associated with several challenges, including non-sustainability, low adherence and inflexibility. AI chatbots are able to overcome these challenges and provide personalized on-demand support, higher interactivity and higher resilience.
AI chatbots use data input from various sources, which is followed by data analysis, which is done through natural language processing (NLP) and machine learning (ML). The output data then helps users achieve their health behavior goals.
In this way, AI chatbots are capable of promoting various health behaviors through the effective delivery of interventions. Additionally, this technology can provide additional benefits for health behavior changes through integration into embodied functions.
Most previous research conducted on AI chatbots aimed to improve mental health outcomes. Relatively recent studies have increasingly focused on the use of AI chatbots to drive health behavior changes.
However, one systematic review of the impact of AI chatbots on lifestyle change was associated with several limitations. These include the authors’ inability to distinguish AI chatbots from other chatbots. Furthermore, this study only targets a limited set of behaviors and does not discuss all potential platforms that could use AI chatbots.
A new systematic review published on the preprint server medRxiv* discusses the results of previous studies of AI chatbot intervention features, functionality, and components, as well as their impact on a wide range of health behaviors.
About the research
The current study was conducted in June 2022 and followed PRISMA guidelines. Here, three authors searched seven bibliographic databases, including IEEE Xplore, PubMed, JMIR Publications, EMBASE, ACM Digital Library, Web of Science, and PsychINFO.
The search involves a combination of keywords that belong to three categories. The first category includes keywords that are related to AI-based chatbots, the second includes keywords related to health behaviors, and the third focuses on interventions.
The search inclusion criteria were studies that included intervention studies focused on health behaviors, those that were developed on existing AI platforms or AI algorithms, empirical studies that used chatbots, English articles published between 1980 and 2022, as well and studies that reported quantitative or qualitative outcomes of the intervention. All data were extracted from these studies and were quality assessed according to the National Institutes of Health (NIH) quality assessment tool.
A total of 15 studies met the inclusion criteria, most of which were distributed in developed countries. The average sample size is 116 participants, while the median is 7,200 participants.
Most of the studies included adult participants, while only two included participants under the age of 18. All study participants had preexisting conditions and included individuals with lower physical activity levels, obesity, smokers, substance abusers, breast cancer patients, and Medicare recipients.
Targeted health behaviors include smoking cessation, promoting a healthy lifestyle, reducing substance abuse, and adherence to medication or treatment. Furthermore, only four studies were reported to use randomized control trials (RCTs), while others used a quasi-experimental design.
The risk of outcome reporting and randomization bias was low, the risk of bias from planned interventions was low to moderate, the risk of bias in outcome measurement was moderate, and the risk of outcome bias was high. All factors for the description of the AI components were sufficient, except for the processing of unavailable input data and the characteristics of the input data.
Of the 15 studies, six reported feasibility in terms of the average number of messages exchanged with the chatbot per month and safety. In addition, 11 studies reported usability in terms of content usability, chatbot ease of use, user-initiated conversation, non-judgmental safe space, and out-of-office support. Acceptability and engagement were reported in 12 studies in terms of satisfaction, retention rate, technical issues and length of engagement.
An increase in physical activity was reported in six studies, along with an improvement in diet in three studies through chatbot-based interventions. Smoking cessation was reported in the four studies evaluated, while one study reported a reduction in substance use and two studies reported an increase in treatment or medication adherence through the use of chatbots.
Several behavior change theories have been integrated into chatbots, including the Transtheoretical Model (TTM), Cognitive Behavioral Therapy (CBT), Social Cognitive Theory (SCT), the Habit Formation Model, Motivational Interviewing, Mohr’s Supportive Accountability Model, and Emotional focused therapy to provide motivational support and behavioral monitoring of participants. Most of the studies focused on behavioral goal setting, used behavioral monitoring, and offered behavioral information, while four studies also provided emotional support.
Most studies use various AI techniques such as ML, NLP, Hybrid Health Recommender Systems (HHRS), hybrid techniques (ML and NLP) and face tracking technology to deliver personalized interventions. Chatbots used mostly text-based communication and were either integrated into existing platforms or delivered as independent platforms. In addition, most chatbots require data about users’ basic information, their goals, and feedback on behavioral outcomes to ensure the provision of personalized services.
Taken together, AI chatbots can effectively promote healthy lifestyles, smoking cessation, and adherence to treatment or medication. Additionally, the current study found that AI chatbots demonstrate significant usability, feasibility, and acceptability.
Taken together, AI chatbots are capable of delivering personalized interventions and can be scalable to diverse and large populations. However, further studies are needed to obtain an accurate description of AI-related processes, as AI chatbot interventions are still in their infancy.
The present study did not include a meta-analysis and focused only on three behavioral outcomes. In addition, articles from unselected databases, articles in other languages, gray literature, and unpublished articles were not included in the study.
An additional limitation was that the interventions could not provide a clear description of the excluded AI chatbots. The study is also not generalizable and information on patient safety is limited.
medRxiv publishes preliminary scientific reports that are not peer-reviewed and therefore should not be considered conclusive, guide clinical practice/health-related behavior, or be treated as established information.