This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Mental Health, is properly cited. The complete bibliographic information, a link to the original publication on https://mental.jmir.org/, as well as this copyright and license information must be included.
Loneliness is a growing public health issue that has been exacerbated in vulnerable groups during the COVID-19 pandemic. Computer agents are capable of delivering psychological therapies through the internet; however, there is limited research on their acceptability to date.
The objectives of this study were to evaluate (1) the feasibility and acceptability of a remote loneliness and stress intervention with digital human delivery to at-risk adults and (2) the feasibility of the study methods in preparation for a randomized controlled trial.
A parallel randomized pilot trial with a mixed design was conducted. Participants were adults aged 18 to 69 years with an underlying medical condition or aged 70 years or older with a Mini-Mental State Examination score of >24 (ie, at greater risk of developing severe COVID-19). Participants took part from their place of residence (independent living retirement village, 20; community dwelling, 7; nursing home, 3). Participants were randomly allocated to the intervention or waitlist control group that received the intervention 1 week later. The intervention involved completing cognitive behavioral and positive psychology exercises with a digital human facilitator on a website for at least 15 minutes per day over 1 week. The exercises targeted loneliness, stress, and psychological well-being. Feasibility was evaluated using dropout rates and behavioral observation data. Acceptability was evaluated from behavioral engagement data, the Friendship Questionnaire (adapted), self-report items, and qualitative questions. Psychological measures were administered to evaluate the feasibility of the trial methods and included the UCLA Loneliness Scale, the 4-item Perceived Stress Scale, a 1-item COVID-19 distress measure, the Flourishing Scale, and the Scale of Positive and Negative Experiences.
The study recruited 30 participants (15 per group). Participants were 22 older adults and 8 younger adults with a health condition. Six participants dropped out of the study. Thus, the data of 24 participants were analyzed (intervention group, 12; waitlist group, 12). The digital human intervention and trial methods were generally found to be feasible and acceptable in younger and older adults living independently, based on intervention completion, and behavioral, qualitative, and some self-report data. The intervention and trial methods were less feasible to nursing home residents who required caregiver assistance. Acceptability could be improved with additional content, tailoring to the population, and changes to the digital human’s design.
Digital humans are a promising and novel technological solution for providing at-risk adults with access to remote psychological support during the COVID-19 pandemic. Research should further examine design techniques to improve their acceptability in this application and investigate intervention effectiveness in a randomized controlled trial.
Australia New Zealand Clinical Trials Registry ACTRN12620000786998; https://www.anzctr.org.au/Trial/Registration/TrialReview.aspx?id=380113
This study investigated the feasibility and acceptability of a digital human (DH) that delivered a psychological intervention to mitigate the effects of social restrictions on loneliness, stress, and well-being in vulnerable populations during the COVID-19 pandemic. The results will inform the design of a randomized controlled trial (RCT) to evaluate intervention effectiveness. To provide a rationale and context for the study, the introduction describes the effects of the COVID-19 pandemic on loneliness, the importance of treating loneliness, and previous work on robot and conversational agent (CA) interventions for loneliness.
Many countries have adopted socially restrictive public health measures over recent months to slow the spread of the COVID-19 pandemic, including the United Kingdom, Canada, the European Union, Japan, and Australia [
While these precautions can help protect vulnerable populations, there are mental health implications of strict social distancing, including increased loneliness [
Loneliness is a subjective psychological state in which a person perceives a mismatch between their actual and desired social relations [
Loneliness interventions can be delivered in-person or remotely through technology, and both have been shown to be effective [
Artificial agents may be a particularly engaging way to provide psychological support to people during a pandemic. People have been shown to feel a sense of social presence with artificial agents, which can improve technology engagement [
A recent scoping review on robot-facilitated loneliness interventions found evidence supporting their use with older adults [
Research looking at the clinical effectiveness of CAs in health care is relatively limited, and a more robust methodology is required [
CAs also show promise for delivering psychological therapies to improve stress and well-being; outcomes that may be worsened by chronic loneliness. Vivibot, a Facebook messenger chatbot that delivered positive psychology exercises over 4 weeks, was found to be acceptable and effective for reducing anxiety in young adults with a chronic health condition [
DHs are a new type of CA that use artificial intelligence to build social and emotional engagement with users [
This study aimed to investigate whether a DH was a feasible and acceptable method of delivering a remote loneliness and stress intervention to high-risk adults during the COVID-19 pandemic. In addition, this study evaluated the feasibility of the study methods in advance of a future definitive RCT. It was hypothesized that a DH would be a feasible and acceptable method of intervention delivery, and that the study methods would be feasible. The results will inform the design of an RCT to investigate the effectiveness of the DH intervention.
This trial was reported in keeping with the CONSORT (Consolidated Standards of Reporting Trials) 2010 statement extension for randomized pilot and feasibility trials [
A randomized pilot trial was conducted involving a parallel mixed design with a waitlist control condition (1:1 allocation ratio). The primary outcomes were feasibility and acceptability, and the secondary outcomes were rapport with the DH, loneliness, stress, COVID-19 distress, positive and negative experiences, and psychological well-being. No major changes were made to the methods after commencing the trial.
The DH facilitator (“Bella”) was developed by Soul Machines Ltd (Auckland, New Zealand) (
Bella autonomously responded to participants’ language using a finite state conversation engine with preprogrammed responses. Bella was programmed to have some autonomous variation in her language for phrases that would not affect her intervention delivery (eg, she varied her greetings each day). Bella spoke using a computer-generated female voice with an Australian accent (“Wavenet C – female” by Google). Participants could communicate with Bella in 1 of the following 3 ways: (1) speech, (2) typing, and (3) clicking on-screen buttons (where present). Bella always responded to participants in speech; however, if participants opened the messenger window to type, they could see a typed version of Bella’s speech as well (see
The digital human's user interface when using the messenger function.
Bella engaged in human-like facial and body gestures as she spoke, including blinking, maintaining eye gaze, raising her eyebrows, and moving her head and shoulders. She showed emotional expressions on her face as she spoke to portray joy and concern, which were preprogrammed and triggered by phrases she spoke using text-to-speech emotional markup language. This involved a process of manually tagging language in her script to elicit particular facial emotions each time Bella spoke the phrase. Bella’s facial expressions were autonomously generated in real-time using visual computing and neurobehavioral modeling techniques (described in greater detail in previous reports [
Bella was designed to deliver several relationship building strategies derived from psychology [
Participants were informed that Bella continuously collected speech and video data in order to communicate (eg, to hear speech and to make eye contact). These data were not recorded, saved, or analyzed by the researchers. Bella’s data collection and use processes are in keeping with the European Union General Data Protection Regulation (GDPR) [
Participants were asked to prioritize visiting the mental health and well-being content that Bella offered as part of their daily website visit. This content included evidence-based exercises to improve loneliness, stress, and psychological well-being, as described below.
Three challenges were delivered over 3 days and included evidence-based strategies to improve loneliness and psychological well-being. The first 2 challenges were (1) to make contact with an old friend, relative, or someone the participant had not been in touch with for a while and (2) to contact someone to let them know something that the participant appreciated about them. These tasks aimed to increase opportunities for social interaction, strengthen social support, and improve social skills. The third challenge asked the participant to make a list of 3 things that they were grateful for, as a positive psychology exercise. Each of the challenges was accompanied by examples to help the participant generate ideas (eg, on day 2, Bella told participants something that she appreciated about them). At the end of the module, participants were reminded to continue practicing kindness toward others and themselves.
This module provided psychoeducation about stress and stress awareness through verbal explanations and diagrams over 1 visit (
The digital human interface during psychoeducation as part of the brain and stress module.
Six modules each focused on a separate psychological well-being tip. The tips encouraged social connection, exercise, acknowledging feelings, being mindful of anxiety-provoking news media consumption, doing activities that elicit positive emotions, and trying out behaviors from a self-care guide.
Participants were able to talk with Bella about a range of other topics beyond mental health and well-being. This included information about the COVID-19 pandemic (eg, New Zealand’s alert levels, details about the virus, symptoms and prevention, and New Zealand’s Healthline and health support resources), and information regarding business and entrepreneurship (eg, remote work and business support organizations).
Thirty participants were recruited. Participants were adults who were at greater risk of developing severe illness if they contracted COVID-19, and as a result, they were asked by the local New Zealand Government to self-isolate to a greater degree during the pandemic. They included (1) older adults aged 70 years or older and (2) adults aged 18 to 69 years who had at least one underlying medical condition that increased the risk of contracting severe COVID-19. The underlying medical condition could have included a serious respiratory disease (such as a chronic lung disease or moderate to severe asthma), a serious heart condition, an immunocompromised condition (such as cancer treatment, smoking-related illness, bone marrow or organ transplantation, hematologic neoplasms, immune deficiency, uncontrolled HIV or AIDS, and prolonged use of corticosteroids and/or other immune-weakening medications such as disease-modifying antirheumatic drugs), a BMI of 40 or higher, diabetes, chronic kidney disease, dialysis, liver disease, and/or pregnancy at the third trimester stage. Participants were required to have English fluency, and access to a computer and internet connection at home. Participants who were 70 years or older were required to achieve a score of 25 or higher on the Mini-Mental State Examination (MMSE). Potential participants were excluded if they received a score of 24 or lower on the MMSE, which would indicate cognitive decline to a moderate or greater degree. Participants aged 70 years or older were not excluded on the basis of whether or not they had an underlying health condition, as their age placed them at a higher risk of developing severe COVID-19.
Twenty-two older adult participants (aged 70 years or older) were recruited from 5 Summerset retirement village sites around the greater Auckland area. Recruitment methods involved presentations to residents about the research, email flyers, and caregiver word of mouth. Residents approached the research team if they were interested in participating. Eligibility screening involving the MMSE and an informed consent procedure (for those who were eligible) were conducted in-person at the retirement village with a member of the research team.
Eight younger adult participants (aged 18-69 years with an underlying medical condition) were recruited from a flyer posted to a staff email list at the University of Auckland, in addition to targeted Facebook advertising, word of mouth, and a Summerset retirement village presentation. Younger adults interested in taking part completed an eligibility screen and informed consent procedure online via a survey website (Qualtrics), except 1 participant who was recruited from a retirement village presentation. This participant completed an eligibility screen and informed consent procedure in-person.
A recruitment target of 30 participants was set, as a minimum of 12 participants per group is recommended for a feasibility study due to precision about the mean and variance [
Data were collected from online questionnaires using Qualtrics, which participants completed from their place of residence. For older adult participants, this may have included completion from a Summerset retirement village independent living villa or apartment, or from the nursing home facility. For younger adult participants, participation took place online from their place of residence in the community or a Summerset care home facility. Data collection took place between November 16, 2020, and March 11, 2021. All participants in the study were provided with a NZ $30 (US $21.50) shopping voucher for their involvement in the research.
Participants were randomly allocated to an intervention or waitlist control group by a member of the research team (EB) (1:1 allocation ratio). Simple randomization was performed using a computerized sequence generation software called Research Randomizer. Allocations were concealed in sealed opaque envelopes from the researchers who enrolled participants (KL and IP) until after participants were enrolled and allocated an ID code. At this point, the researcher was deblinded to assign participants to conditions and provide participants with the appropriate instructions. Participants were deblinded after their assignment to conditions.
Once enrolled, participants were contacted over email with instructions for proceeding in the trial. For nursing home residents, their caregiver was copied in the email communications and facilitated the participant’s involvement in the study.
All participants completed an online baseline questionnaire on day 1 of their participation. Then, participants in the intervention group completed a DH training session with a member of the research team for 30 minutes. For all older adults (plus 1 younger adult participant), this took place in-person at their retirement village or nursing home facility. For 7 of 8 younger adults, this took place either in-person at the University of Auckland Clinical Research Centre or online over Zoom video conferencing software (Zoom Video Communications), depending on the lockdown conditions.
All participants received the same technology training, which involved learning how to interact with Bella and completing “day 1” of their intervention week with the researcher present to answer questions. The researcher ensured that the software worked on each participant’s computer. Participants were provided with written instructions and pictures of the user interface that summarized the training session content. For sessions over Zoom videoconferencing, the screen share feature was used and participants received a PDF copy of the interaction instructions. Three participants were trained over Zoom, and 23 participants were trained in-person.
Participants were asked to interact with Bella for at least 15 minutes per day over 1 week. Participants visited Bella’s website independently from their place of residence. The daily 15 minutes could include time spent interacting with Bella and doing therapy activities (eg, a deep breathing exercise). They were asked to prioritize completing the mental health and well-being modules before visiting other content. Participants interacted with Bella at their chosen time of day. Participants were sent a daily text reminder to engage in the intervention and were informed that they could text back to receive technical support.
On day 8, the intervention group finished their intervention week and completed an online postintervention questionnaire. One week later, on day 15, intervention group participants filled out an online follow-up questionnaire.
For participants in the waitlist control group, the order of the procedure was slightly different. Participants in the waitlist group completed an online baseline questionnaire on day 1 and then waited for 1 week. On day 8, waitlist participants completed a second online questionnaire, completed the technology training session, and began their intervention week. On day 15, at the end of their intervention week, waitlist participants completed the postintervention questionnaire.
Time points for assessments. PSS-4: Perceived Stress Scale 4 items; SPANE: Scale of Positive and Negative Experiences.
Observations were made by a member of the research team (KL) about how the intervention was used (eg, independently or with the aid of a caregiver), along with dropout rates and reasons. Observations were also made regarding the feasibility of the technology training methods for younger and older adults, and nursing home residents. Instances were recorded where participants refused to receive training through a particular delivery method (eg, video calling).
Observations were recorded during recruitment and data collection by a member of the research team (KL). Observations pertained to the success rate of different recruitment strategies for younger and older adult participants, and challenges associated with data collection from the online forms that participants completed independently.
Bella’s acceptability was measured using quantitative self-report items and open-ended qualitative questions designed for the study. Behavioral engagement data were also collected. The acceptability measures are outlined in further detail below.
Participants were asked to rate whether (1) they felt Bella was helpful for promoting resilience and psychological well-being, (2) they felt Bella was helpful for improving feelings of loneliness, and (3) they would be willing to use Bella again in the future, using a 7-point scale with response anchors (1, “definitely no” to 7, “definitely yes”).
Participants provided written responses to the following open-ended questions:
Behavioral engagement with Bella over 1 week was evaluated by retrospective self-report. Participants reported on which days of the week they visited Bella and estimated approximately how long they used Bella each day in minutes.
The acceptability of each psychological intervention module was evaluated separately in the postintervention questionnaire. Participants rated how much they liked the brain and stress module, Headspace (if they visited), and the expressing kindness challenge (including each of its 3 activities) on a 7-point scale with response anchors (1, “not at all” to 7, “very much”). Participants rated how beneficial they found the expressing kindness challenge for well-being, and how well they felt the brain and stress module improved their understanding of the stress response on a 7-point scale (1, “not at all” to 7, “very much”). Participants who visited Headspace were asked whether they felt that Headspace was a helpful resource to link to with a dichotomous yes/no response option. The participants were also asked the following qualitative question:
Rapport with Bella was measured using the 20-item Friendship Questionnaire developed by Johanson et al [
Loneliness was measured using the 20-item UCLA Loneliness Scale (Version 3) [
Perceived stress was measured using the 4-item Perceived Stress Scale (PSS-4) [
Worry about contracting COVID-19 was measured using a 1-item scale [
The Scale of Positive and Negative Experiences (SPANE) has two 6-item subscales that measure positive emotions (SPANE-P) and negative emotions (SPANE-N) [
Psychological well-being was measured using the 8-item Flourishing Scale [
Data were analyzed using SPSS software (version 27; IBM Corp). Missing data were addressed by imputing the mean score of the participant’s other responses to the scale at the timepoint. For 1-item scales, where it was not possible to impute a score or where the participant did not complete a full scale, the participant’s data were excluded from analysis of the relevant variable.
Baseline demographic and psychological variables were calculated for the overall sample, and compared between groups using chi-square tests and independent samples
Written responses to 3 open-ended questions were analyzed using reflexive thematic analysis [
Participants were predominantly female (24/30, 80%) and Caucasian (22/30, 73%), and mainly had high school or less education (14/30, 47%). Half of the sample (15/30, 50%) reported an underlying medical condition. Participants reported low levels of loneliness (mean 37.79, SD 9.90) and stress at baseline (mean 3.86, SD 2.88). Participant characteristics at baseline are reported in
CONSORT (Consolidated Standards of Reporting Trials) diagram of participant flow.
Older adults required technology training to be completed in-person at their retirement village with a member of the research team (KL or IP). This method worked well as it avoided any discomfort with using the video calling software. One older adult was offered training over the Zoom video calling software during the lockdown period and refused as she was not able to use the software. Only 1 of 22 older adult participants did not have a webcam as part of their computer, which was uncovered during the technology training (the DH website requires a webcam). To solve this, a webcam was borrowed from the retirement village reception and installed by a member of the research team (IP) during the training session.
Younger adult participants were generally able to be trained either in-person from a clinic room at the university (outside of the lockdown period) or online over Zoom (during the lockdown period). Video calling did not appear to impact the effectiveness of the training. Technical support requests were low for younger adults during the study, irrespective of how their training was delivered. One younger adult participant who was a nursing home resident required in-person technology training.
Six participants withdrew from the study (all older adults). The reasons for withdrawal were as follows: (1) the Wi-Fi speed at the retirement village location was too slow for Bella to load properly (n=2); (2) cognitive health difficulties interfered with understanding study instructions (n=1); (3) the participant was too busy to take part after enrollment (n=2); and (4) technology training could not be scheduled (n=1).
Twenty-four participants completed the intervention, 22 of whom completed it independently after training. Two participants (1 younger adult and 1 older adult) who were both nursing home residents required caregiver assistance to access the website and interact with Bella.
The majority of older adult participants were recruited through information sessions held at retirement villages (21/22, 96%). One older adult participant was recruited through an email flyer sent by a village staff member to residents.
For younger adults, the most effective recruitment method was by advertisement to a university junk email list (5/8, 63%). Facebook advertising and word of mouth each resulted in recruitment of 1 participant, and 1 participant was recruited from an information session at a nursing home.
Most participants completed assessments online using Qualtrics without significant issue. Two participants reported instances where they were unsure whether their responses had been submitted. Nursing home residents were unable to complete assessments independently on a website and required caregiver assistance.
Overall, participants reported that Bella was somewhat helpful for promoting resilience and psychological well-being (mean score 4.39 out of 7, SD 1.83) and for improving any feelings of loneliness (mean score 4.09 out of 7, SD 1.76), as responses on average were above the mid-point. Participants were somewhat willing to use Bella again in the future (mean score 4.09 out of 7, SD 1.98). Younger and older adults rated Bella similarly across the acceptability items, and no significant differences were found.
On average, participants interacted with Bella 6 out of 7 days (mean 6.23, SD 1.19). Participants interacted with Bella for approximately 20 minutes per day (mean 20.20, SD 13.95); 5 minutes longer than the 15 minutes per day requested by the researchers. The average total interaction time with Bella over 1 week was 128 minutes (mean 128.33, SD 102.77). There were no significant differences between younger and older adults in engagement behavior.
Participants identified several strengths and limitations of Bella through responses to the following 2 written open-ended questions:
Themes, subthemes, and representative quotes describing what participants liked most about Bella.
Themes and subthemes | Representative quotes | |
|
|
|
|
Facial expressions | |
|
Human-like | |
|
Attractive face | |
|
Similarity to user | |
|
|
|
|
Gentle voice | |
|
Self-disclosure | |
|
Clear language | |
|
|
|
|
Companionship | |
|
Nonjudgmental | |
|
Friendly personality | |
|
Validating | |
|
Calm personality | |
|
|
|
|
Quality resources | |
|
Accessible delivery | |
|
|
|
|
Interaction modalities | |
|
User controls interaction | |
|
Easy to use | |
|
Overall experience | |
|
|
|
|
Something different | |
|
A new technology |
Themes, subthemes, and representative quotes describing what participants felt could be improved about Bella.
Themes and subthemes | Representative quotes | |
|
|
|
|
More human-like movements | |
|
Ability to touch | |
|
More positive emotional expression | |
|
|
|
|
Extra conversation topics | |
|
Greater interactivity | |
|
More personalized responses | |
|
Regularly update information | |
|
Avoid human-like backstories | |
|
Address user by name | |
|
Incorporate user’s life experience | |
|
|
|
|
A more human-like voice | |
|
Formal speech delivery | |
|
|
|
|
Natural language understanding | |
|
General technology advances | |
Preference for a real human | ||
No changes |
Of 24 participants, 22 visited the expressing kindness challenge. Overall, the expressing kindness challenge was liked by participants (mean score 5.50 out of 7, SD 1.34), as were the 3 daily challenges of which it was comprised. Participants reported liking reaching out to a friend (mean score 5.95 out of 7, SD 1.13), telling a friend what they appreciate about them (mean score 5.77 out of 7, SD 1.09), and making a gratitude list (mean score 5.71 out of 7, SD 1.14). Participants reported that the expressing kindness challenge felt beneficial for their well-being (mean score 5.00 out of 7, SD 1.95). There were no significant differences between younger and older adults in terms of how much they reported to like the expressing kindness challenge (mean score 5.57, SD 1.39 vs mean score 5.47, SD 1.36;
The majority of participants (13/24, 59%) visited all 3 tasks of the expressing kindness challenge. Two participants (9%) visited only 2 tasks, and 7 participants (32%) visited only 1 task. Most participants (15/24, 68%) completed the expressing kindness challenge on 3 consecutive days. One participant completed the challenge in 1 day, and 6 participants (27%) completed the challenge in other ways (eg, spread over a week).
Most participants who visited the expressing kindness challenge attempted the activities. All 20 participants who visited day 1 completed the activity (ie, reaching out to a friend). Of 16 participants who visited day 2, 15 completed the activity (ie, telling a friend what they appreciate about them). All 14 participants who visited day 3 did the activity (ie, make a gratitude list).
Twenty-one participants visited the brain and stress module. On average, participants reported that they liked the brain and stress module (mean score 5.52 out of 7, SD 1.25), and that it improved their understanding of the stress response (mean score 4.90 out of 7, SD 1.61). There were no significant differences in how much younger and older adults liked the brain and stress module (mean score 5.71, SD 1.38 vs mean score 5.43, SD 1.22;
Of 21 participants who visited the brain and stress module, 18 (86%) reported looking at the mental health tips section afterwards to learn about stress management and mental well-being. Additionally, 17 participants (81%) visited Headspace, which is a meditation website that the DH linked to at the end of the brain and stress module [
Participants visited an average of 9.39 (SD 5.23) other modules beyond the expressing kindness challenge and the brain and stress module (ie, the mental health modules that the researchers asked them to complete in particular). There were no significant differences in how many additional modules younger and older adults visited (mean 8.50, SD 6.37 vs mean 9.87, SD 4.69;
Seventeen participants responded to the question
Overall, participants reported a reasonable degree of rapport with Bella (mean score 66.92 out of 100, SD 12.63). There was no significant difference in the amount of rapport reported by younger and older adults (mean 68.13, SD 14.86 vs mean 66.31, SD 12.45;
There was no significant main effect of time (
There was a significant main effect of condition on perceived stress (
There was no significant main effect of time (
There was no significant main effect of time (
There was no significant main effect of time (
There was no significant main effect of time (
There was a trend toward a significant main effect of condition on psychological well-being (
Technology has come to play an important role in combatting the COVID-19 pandemic. Artificial intelligence technologies have been rapidly deployed to assist in diagnosing COVID-19 cases and forecasting epidemic development, contact tracing, aiding in drug and vaccine discovery research, and predicting patient outcomes such as disease severity, length of hospital stay, and mortality risk [
This study found that a DH was a feasible and acceptable way to deliver a remote loneliness and stress intervention to at-risk older adults living independently and to younger adults with a chronic health condition based on behavioral, qualitative, and some self-report data. The intervention was less feasible for nursing home residents who required caregiver assistance to participate, which may have increased caregiver burden.
Prior to the pandemic, evidence had been building in support of the effectiveness, feasibility, and acceptability of CAs, including embodied agents, at delivering remote psychology interventions and assessments [
A challenge of evaluating the effectiveness of CAs in psychology applications is the large heterogeneity of outcome measures, psychological interventions, and technology features across the literature, which makes comparisons difficult, alongside a shortage of RCTs [
Questions remain around how to optimally design CAs for health care applications [
This study investigated a novel application of DH technology and adopted a pilot RCT design to inform the methodology of future trials. However, there were several methodological limitations. A sample bias may have occurred whereby participants who volunteered may have been more digitally literate or comfortable with using novel technologies. Moreover, the sample predominantly included Caucasian women; therefore, it is unclear how well the results would generalize to a more diverse population. Even though randomization was conducted, there were significant group differences at baseline in stress, and a larger sample likely would have eliminated these differences. Changes in and out of lockdown conditions in Auckland during the data collection period could have affected the psychological results and degree of engagement in the study. Moreover, there was no control for the psychological follow-up data of our waitlist group, and this should be addressed in a future trial. It is also unclear what the level of engagement with the DH would be outside of a clinical trial context. Research has shown that engagement with eHealth interventions is often lower than what is observed in trials [
The results suggest several directions for future research. A fully powered RCT should investigate the effects of the DH intervention on loneliness and stress. This trial could address the methodological limitations of our study. An active control condition (eg, a chatbot and a website) could be used to provide stronger evidence of effectiveness and reduce the chance that outcome improvements are due to confounding variables (eg, passage of time and researcher attention). The length of follow-up for psychological measures should be extended, along with the length of the intervention. Many loneliness interventions take place over 4 to 6 months with weekly sessions that take an hour or more [
Bella, a DH, was found to be a feasible and acceptable way to deliver a remote loneliness intervention to at-risk adults facing social restrictions during the COVID-19 pandemic, based on behavioral, qualitative, and some self-report data. The results support conducting a larger and longer RCT to investigate intervention effectiveness, and indicate that several changes should be made to the technology, intervention content, and trial design. DHs are a novel technological solution that may provide remote psychological support to socially restricted at-risk groups during pandemics. Research should examine the use of DHs in other health care applications with diverse patient populations.
The Friendship Questionnaire (Johanson et al, 2020 [<xref ref-type="bibr" rid="ref53">53</xref>]) adapted to the digital human.
Participant characteristics at baseline.
Definitions of themes in response to the qualitative questions.
Number of participants who visited each module.
Conversation topics that participants would like to talk about with Bella.
Perceived stress (mean scores) between groups over the 3 time points.
Psychological well-being (mean scores) between groups across the 3 time points.
CONSORT-eHEALTH checklist (V 1.6.1).
conversational agent
digital human
Mini-Mental State Examination
Perceived Stress Scale
randomized controlled trial
Scale of Positive and Negative Experiences
We sincerely thank Summerset retirement village and the staff for supporting the research. Orquidea Tamayo Mortera, Steve Williamson, Bobby Gerneke, Claire-Marie Blair, Mike Wharepouri, Jo Clark, Abba Samman, Joseph Lascano, Lianne Gray, and Kate Chastukhina played an integral part in the success of our recruitment of residents and their involvement in the research throughout. We also wish to acknowledge the engineering team at Soul Machines Ltd for their support with developing the digital human. In particular, we would like to thank Erica Lloyd, Hilary O’Connor, Zara Jillings, Keegan Swenson, and Nitin Bhandari. We also wish to thank Roisin Kelly for contributing illustrations for the brain and stress module. We further thank Mikaela Law and Gabrielle Sebaratnam for their assistance with conducting eligibility screens and enrolling participants at Summerset retirement village. Soul Machines Ltd supported the research by building the digital human.
MS is the chief executive officer of Soul Machines Ltd (a New Zealand artificial intelligence company), which supports KL with a PhD stipend and contracts EB for consultancy work.