Enhancing Self-Efficacy in Health Self-Examination through Conversational Agent's Encouragement

Research Background and Issues

• Identified Problems or Challenges: This paper explores the primary barriers people face when conducting health self-checks (e.g., skin cancer risk assessments), including a lack of confidence or knowledge. These barriers may lead to delayed medical consultation or overconfidence resulting in misjudgment. Additionally, while traditional medical fields emphasize the use of self-efficacy to enhance individual behavior, research on AI systems supporting health behaviors remains limited.
• Why It Matters: Early detection of skin cancer, particularly deadly melanoma, is critical in reducing mortality rates. The positive correlation between self-checks and health outcomes has been confirmed, yet many neglect this behavior due to a lack of knowledge and confidence. Supporting self-checks and reducing medical delays can save lives.
• Research Motivation and Related Work:
  • Research Motivation: To design an AI conversational agent (CA) that uses encouraging language to influence users' self-efficacy and trust, aiming to help users conduct health self-checks with greater confidence.
  • Related Work: Previous studies have shown that encouraging language and early successful experiences are effective in enhancing self-efficacy. Moreover, patient-centered physician behavior has been proven in prior research to significantly improve patient trust and health behaviors.

Solution

• Method or Solution:

  • The authors designed an AI conversational agent (CA) capable of providing two types of language feedback—encouraging feedback and neutral feedback.
  • Using a 2×2 experimental design (encouraging/neutral language and success/failure mastery experiences), the study examines the impact of CA language type on users' self-efficacy and trust.

• Innovations:

  • This study integrates self-efficacy theories (e.g., Bandura's Social Cognitive Theory) with AI technology, endowing the AI with personalized characteristics to provide emotional support through voice feedback.
  • It balances key dimensions of "trust": competence, benevolence, and integrity, specifically exploring how encouraging language affects users' perceived benevolence and self-efficacy.

• Implementation Steps and Key Technologies:

  1. Conversational Agent Design: Develop a voice-based CA that offers two types of language behaviors (encouraging and neutral).
     • The encouraging agent enhances user confidence through positive affirmations (e.g., "You can absolutely do this!").
     • The neutral agent maintains emotional neutrality, focusing on information delivery.
  2. Experimental Method:
     • Participants were randomly assigned to four experimental conditions (encouraging language + success experience; neutral language + success experience; encouraging language + failure experience; neutral language + failure experience).
     • Each participant completed multiple skin self-check tasks and received real-time feedback from the CA.
  3. Measurement Metrics:
     • Self-efficacy was measured through surveys at different stages of the task (pre-interaction, during interaction, post-interaction).
     • User trust was quantified based on scores across three dimensions: competence, benevolence, and integrity.
  4. Technical Implementation:
     • Feedback was generated using the OpenAI GPT model.
     • Google Speech API was integrated to enable voice interaction.

Research Outcomes

• Specific Findings:

  • Encouraging language significantly improved users' self-efficacy, with self-efficacy scores during interaction increasing notably (average score rose from 3.92 to 4.11), and higher trust scores in the benevolence dimension.
  • Participants reported in qualitative feedback that the CA's combination of voice and imagery enhanced learning outcomes and made the learning process more personalized.

• Comparison with Existing Solutions and Advantages:

  • Compared to text-based manuals or traditional learning materials, the CA's interactive format was considered more engaging and better at fostering participation.
  • Encouraging language was perceived by participants as more effective than traditional neutral language in boosting confidence and aligning with a more humanized communication style.

• Experimental or Evaluation Results:

  • The group with the encouraging CA scored higher in trust compared to the neutral CA group (especially in the benevolence dimension, p = 0.019).
  • The experiment did not find a significant impact of language type or mastery experience on task completion accuracy, indicating that increased confidence does not directly translate to immediate performance improvement.

• Limitations and Future Directions:

  1. Short-Term Effect Limitation: The positive effects of encouraging language may only be effective during the interaction and may not contribute to long-term efficacy improvement.
  2. Single Task Domain Limitation: The study focused solely on skin self-checks, which may not directly apply to other health behavior domains.
  3. Real-World Validation: Conducted in a lab environment, further research is needed to determine whether the findings can effectively support users in real-world settings.
  4. Future Directions:
     • Investigate changes in self-efficacy during long-term interactions with the CA.
     • Expand research to other domains (e.g., medication adherence or mental health support).
     • Explore the dynamic mechanisms of trust-building to understand how trust evolves over time during interactions.

Through this study, the authors provide critical insights into designing conversational agents to better support users in conducting health self-checks, while also cautioning against the risks of overconfidence resulting from enhanced self-efficacy. This work offers valuable references for the design and implementation of technology in related health domains.