Evaluating the feasibility, effectiveness, and acceptability of AI-chatbot support in an online intervention for female sexual dysfunction
Distressing low sexual desire is the most prevalent sexual dysfunction among women. One treatment option is eSense, an effective digital tool delivering mindfulness-based therapy (MBT). It integrates human “navigators” to support users and, though they improve eSense adherence, may be costly during scale-up. Large language model (LLM)-based chatbots, or AI-chatbots, are a promising substitute, with research suggesting user-customization may enhance interactions. This thesis explored whether an GPT-4o-based AI voice chatbot could be a feasible and effective substitute for human guidance. In Phase 1, three groups of women (n = 3 per group) with distressing low sexual desire and/or arousal used eSense-MBT alongside the AI-chatbot. Interviews were conducted to gather feedback about the AI-chatbot and understand facilitators/barriers to usage. This was used to improve the chatbot after every group, resulting in three rounds of refinement. Quantitative outcomes of sexual distress and desire/arousal were also collected. In Phase 2, women with distressing low sexual desire and/or arousal were randomly assigned to use eSense-MBT alongside a customizable AI-chatbot or a human navigator (n = 61 per arm). Homework engagement was the primary outcome for non-inferiority analysis and measured after each eSense module. We also measured secondary outcomes of sexual distress, sexual desire/arousal, number of homework assignments completed, treatment satisfaction, and working alliance. Phase 1 identified nine themes across three research questions focused on improvements, user experience, and feasibility. Key themes included: (1) Safety and confidentiality; (2) Expectations of technology; and (3) Comparison to humans. In Phase 2, the chatbot’s performance was not consistently non-inferior to that of the human, suggesting lower homework engagement in the AI-chatbot navigator group. Secondary findings revealed significant improvements in sexual distress and desire/arousal with no differences between navigator arms, more homework completed in the human navigator group, higher working alliance with the human than with the AI-navigator, and a decrease in positive attitudes towards AI post-treatment. Overall, these findings suggest that an AI-chatbot can fulfill navigator responsibilities but does not provide support for homework engagement comparable to a human. Additional considerations for future research on intervention support should include addressing user preferences, technical challenges, and program adherence.