Our study report is guided by the Transparent Reporting of Evaluations with Nonrandomized Designs (TREND) reporting structure for nonrandomized studies [14]. The study was a nonrandomized single arm pre-post study. Study participants conducted a single online study session, involving pretest questionnaires, an interaction with the study intervention, GRACE, and post-test questionnaires.

All participants were recruited online through the Prolific web-based research platform [15]. The inclusion criteria were the following: (1) age 18 years or older; (2) reside in the United States; (3) speak and read English; and (4) have a current or prior diagnosis of cancer. Prolific’s system prescreens participants and handles all aspects of recruitment automatically, offering enrollment to eligible participants until the recruitment target is met. A total of 30 participants who met the inclusion criteria completed the study. We used four “attention check” questions in the surveys, as is common practice in web-based studies [16-18] but no participants failed more than one; so all participants who completed the study tasks were retained. Participants engaged with the intervention remotely. The intervention and measure collection was completely automated.

Ethics approval was received from the Institutional Review Board at Northeastern University (reference 25-04-01). All participants completed an informed consent process, which allowed them to opt out of any survey questions or withdraw from the study at any time. The participants were compensated US $15 after completing the survey. All collected data were deidentified.

We developed an online system called “GRACE” (Genetic Relational Agent for Cascade) that features a relational agent (RA) that interacts with patients with cancer using simulated face-to-face conversation (Figure 1). GRACE was deployed as a web-based application developed using the Unity 3D game engine. The RA uses synthetic speech and animation to simulate a face-to-face conversation, and the patient responds with a touch screen or mouse. All dialogue is scripted so that accuracy and safety can be reviewed. This interface has been successfully used in several health interventions with thousands of patients across the health literacy spectrum [19-24], including a pilot study of a RA that provided pretest genetic testing to patients with cancer [25]. The content of GRACE was developed based on previous work within cancer genetics and cancer genetic testing, whose input was provided by health providers, patients, and families as the content was developed. All clinical content was created and approved by experts in cancer genetic counseling and testing. Previous work has verified that the structure, approach, and style of the Relational Agent is acceptable to patients. It is intended that this content take as much or as little time for the patient to review as necessary; however, it is designed to be brief, less than 20 minutes, and can be administered remotely.

The RA first spends several minutes reviewing the basics of cancer genetics, including the concepts of genes, pathogenic variants, inherited risk, what individuals can do to reduce their cancer risk if they know they have a pathogenic variant, and what the genetic testing process involves. It then elicits from the proband the most likely family member they would talk to first about genetic testing. It then covers a series of modules on communication skills, following Buckman’s six-step strategy [16] augmented with techniques from Motivational Interviewing [17], starting with setting the context, building rapport, and asking permission to discuss cancer genetic testing, before progressing to asking open-ended questions, active listening, and relaying their own story about reasons for getting tested and what the process was like for them. For each module, GRACE first provides didactic instruction on the skill, including examples, before allowing the user to practice each skill in a role-play with a different conversational agent that takes on the persona of the family member the proband indicated they would talk to first. At each turn of the role-play, the proband is given the option of choosing from among one correct response and several alternatives representing common communication mistakes. If the user chooses the correct options throughout the role-play, the instructional RA gives positive reinforcement. However, if they make a mistake, the role-play agent provides immediate in situ feedback (eg, acting offended and leaving), the role-play is ended, and the user is given the option of repeating.

Web-based questionnaires were distributed in English to the participants using a Qualtrics web survey.

Given the exploratory nature of the study, no formal sample size calculation was conducted. The research data were analyzed using IBM SPSS Statistics (version 28, IBM Corp). The distribution of pre-post differences in KnowGene scores was normal, indicating parametric paired sample t-test use for pre-post tests. All pre-post tests on single scale item Family Communication Measures were tested using non-parametric Wilcoxon Signed Ranks Tests. The distribution of scores for the composite measure of satisfaction as well as intervention duration were non-normal, thus non-parametric tests were used. Descriptive statistics were used to summarize the data. The threshold for statistical significance was set at P<.05.

Table 3 shows the sociodemographic characteristics of the 30 study participants. The age of the participants ranged from 19‐79 years (mean 44.9, SD 16.4 y). A majority of participants identified as male (16/30, 53.3%), were White (24/30, 80.0%), and had at least some college education (25/30, 83.3%). The majority had received prior cancer-related genetic testing (17/30, 56.7%).

Following the interaction with GRACE, participants indicated they were significantly more likely to share genetic test results with their family (Table 1). Participants also indicated that the importance of sharing genetic test results was significantly greater following the intervention, and that they felt significantly more confident in discussing genetic test results with their family. There was only a trending increase in the degree of comfort participants felt in discussing genetic test results with their family, P=.07.

KnowGene scores increased from pre-intervention, 13.27 (SD 1.91), to post-intervention, 13.70 (SD .35) but this difference was not significant, paired t₂₉=1.25, P=.22.

Participants scored GRACE an average of 42.7/60 (71.2%) on the SUS.

Participants rated their satisfaction with GRACE an average of 6.20 (SD 0.90) on the composite measure of satisfaction (Table 2). A single sample Wilcoxon signed ranks test indicated this score (mean 6.2) was significantly greater than a neutral score of 4.0, P<.001. There was a significant positive correlation between time to complete the intervention and satisfaction, Spearman ρ=.537, P=.002.

We aimed to understand usability and the preliminary impact of a digital health intervention to support patients with cancer in communicating genetic testing information to at-risk relatives. Overall, we found that comfort, confidence, and intent to communicate increased, and that participants had high levels of satisfaction with the intervention. Cancer genetics knowledge remained constant. Preliminary results demonstrate that the intervention GRACE should be evaluated further to understand the impact on rates of cascade testing.

Our study had some limitations. First, our sample was a convenience sample of individuals recruited from a research database. We did not select people who had completed cancer genetic testing and who had actual testing information to communicate. We therefore could not measure actual communication or uptake within a family. Additionally, the sample is small and relatively homogeneous, recruited from one online crowdsourcing site; therefore, results may vary as the sample diversifies. Specifically, KnowGene scores were relatively high at baseline, and therefore no change was found. If we targeted patients with varied educational levels and knowledge, we may have noticed an impact. This may have contributed to our lack in change in knowledge scores.

Our study is the first to use a conversational agent to provide a fully virtual and automated intervention to promote family communication between a patient with cancer and at-risk relatives. Other forms of digital health interventions have been tested to improve these outcomes. In a recent meta-analysis focused on interventions to improve cascade genetic testing, only two technology-based interventions were identified [11]. One was the Family Gene Toolkit, a webinar-based intervention to promote communication in families with hereditary breast and ovarian cancer syndrome that also included face-to-face and telephone interactions over 5-weeks [7]. The published data for this study demonstrate acceptability [9], however, outcomes have not yet been published. The other was the Mobile Application for Genetic Information on Cancer (mAGIC) trial which focused on ovarian cancer risk communication using a mobile app over 3-months [6]. Outcomes of this pilot randomized trial demonstrated significant increases in hereditary cancer knowledge and family communication rates for the intervention group compared to control, but no significant differences in use of cancer genetic counseling services.

Three additional trials were identified since the publication of the meta-analysis that use technology to promote cascade testing, mainly through the use of video education, which was found to be non-inferior compared to usual care to provide genetics education [29] and also to not cause distress [30]. The Genetic Education, Risk Assessment, and Testing (GENERATE) Study used telemedicine and online genetics education through videos to reach family members of patients with pancreatic cancer and a known pathogenic variant, finding that it was successful at improving uptake of testing [31]. Two protocol papers were available related to video education trials: the Genetic Education for Men (GEM) trial (focused on men with prostate cancer [32]), and the Facilitated Cascade Testing (FaCT) trial [33].

While providing genetics education information through a video can be useful and effective at providing information, there are limitations that can be addressed through more interactive computer programs such as relational agents. Video education is static and often challenging to keep up to date and current, a need for those seeking genetics information. Additionally, providing information passively without interaction diminishes learning outcomes. Video education is also difficult to tailor for the nuances required within a genetics education setting.

Researchers and companies are now developing chatbots to facilitate testing or family communication both focused on cancer and within other inherited domains, such as familial hypercholesterolemia [34-36] . However, chatbots rely entirely on text-based interaction, limiting their use in low-literacy populations. Accessibility by family members across the literacy spectrum is essential to ensure that all at-risk individuals receive the genetic risk and mitigation information they need, including the 46% of US adults with low or marginal health literacy [37]. Additionally, chatbots driven by large language models, can hallucinate and provide false information [38], even when driven by retrieval augmented generation [39], making them potentially unsafe if their output is not entirely screened by clinicians. Therefore, solutions that provide trustworthy, controlled, and precise information are needed to ensure accurate and appropriate information without intensive provider oversight, and in a manner accessible to all.

In order to achieve the full public-health impact of cancer genetic testing, it is critical that information about hereditary cancer risk is communicated within families, often originating from the patient with cancer. The GRACE intervention demonstrates an opportunity to motivate patients with cancer to communicate genetic information and promote cancer genetic testing to at-risk relatives. Further large-scale randomized studies are needed to understand the impact of GRACE on genetic testing outcomes.