Design of Chatbot Persona for Breast Cancer Screening Outreach Among Black Women

Breast cancer screening disparities among Black women persist despite health system recognition and outreach. However, current evidence on how to tailor and optimize implementation strategies for breast cancer screening outreach is limited. The proposed study is part of a larger project to design a chatbot for breast cancer screening outreach to Black women and will focus on optimizing the chatbot persona. Using the Multiphase Optimization Strategy (MOST) framework, the investigators will conduct a randomized factorial experiment to assess the individual components of chatbot persona for breast cancer screening and identify which components have the greatest effect on trust and engagement for Black women. This information will guide the design of an optimized chatbot intervention that achieves the primary outcomes.

The goal of this study is to determine the optimal delivery of initial chatbot messages for culturally tailored breast cancer screening outreach. Mistrust of the medical system has been identified as a significant barrier to mammography screening among Black women. Yet, while tailored interventions for breast cancer screening exist, the optimal design of a tailored intervention to engender trust is unknown. Chatbots have been shown to increase levels of trust in web-based information, though adoption of chatbots may depend on chatbot characteristics. The investigators propose to use the Multiphase Optimization Strategy (MOST), a framework for developing efficacious, efficient, scalable and cost-effective interventions, to assess the performance of chatbot intervention components and their interactions. The chatbot message delivery will be systematically varied across two components, each of which is represented by a separate factor in the 2x2x1 factorial study design with a control arm. Specifically, each participant will be randomly assigned to one of five separate experimental conditions. Conditions include: (1) chatbot with a primary care doctor persona and direct communication style; (2) chatbot with a breast cancer survivor persona and direct communication style; (3) chatbot with a primary care doctor persona and indirect communication style; and (4) chatbot with a breast cancer survivor persona and indirect communication style. All participants will complete a survey regarding their perceptions about the initial outreach messages from the chatbot. The main effects will be estimated of the two experimental factors and their interactions on the study's primary outcomes - trust in the chatbot system to use for breast cancer screening education and scheduling, and intention to use. This information will guide the design of an optimized chatbot persona that achieves the primary outcomes. Participants will be enrolled if they are Black individuals who qualify for breast cancer screening residing in the United States who are between the ages of 40-74. Recruitment will be conducted on Prolific, an online participant pooling platform, and Amazon Mechanical Turk (MTurk), a crowdsourcing platform used for research recruitment. Prolific will be used given the platform's ability for selecting the participant population. However, due to the limited number of individuals within the inclusion criteria on Prolific, and if needed participants will also be recruited on MTurk. Participants will be asked to view the chatbot messages and respond to questions to assess trust, engagement, and directness of the chatbot.

The chatbot messages are characterized by subjunctive modal verb forms (''would like'') and cooperative addresses (''we", "let's").

The human-computer trust scale assesses user trust, which is based on similar constructs of trust (benevolence, competence, reciprocity, perceived risk). 7 of the 12 items were selected which use a 5-point Likert scale from 'Strongly disagree' to 'Strongly agree'.

This measure assesses likelihood to use this system to schedule a mammogram in the future, and is scored on a 5-point Likert scale from 'Very unlikely' to 'Very likely'.

This measure consists of 4 semantic differential scales assessing traits (important, interesting, relevant, warm) on a 7-point scale.

This measure consists of 7 semantic differential scales assessing traits (direct, friendly, caring, straightforward, demanding, respectful, polite) on a 7-point scale.

These 4 items measure the perceived expertise and attitude of the system on a 5-point Likert scale from 'Strongly disagree' to 'Strongly agree'.

This measure consists of 3 items to assess self-brand connection on a 5-point Likert scale from 'Strongly disagree' to 'Strongly agree'.

Inclusion Criteria: All women who are 40-74 years old: who identify as Black Exclusion Criteria: Participants who do not complete the survey Participants who complete the survey in less than half the normal average time Participants who do not pass the attention check

Black or African American women between the ages of 40-74 years old and residing in the United States.