Uncertainty-Aware Health Coaching for Sustainable Habit Building with Human Oversight
Authors: Iva Bojic, Michael Tanzer, Mahnoosh Mehrabani, Ali Dadgar, Srinivas Bangalore, Andy Khong.
Health coaching represents a shift from episodic, clinician-centered care to continuous, personalized support that empowers individuals to build and sustain healthy habits in daily life. As a non-clinical practice, it focuses on goal setting, self-reflection, motivation, and lifestyle behavior change rather than diagnosis or treatment. Evidence shows it effectively improves health behaviors, supports lifestyle modification, and enhances goal attainment; however, its reliance on sustained human involvement makes it labor-intensive and difficult to scale while maintaining consistency and long-term engagement. These limitations motivate the use of AI agents to help scale health coaching and make it more widely available and affordable while preserving its core strengths.
We propose to develop an uncertainty-aware health coaching system that advances safety and reliability through structured human–AI collaboration. The proposed multi-agent system incorporates both real-time intervention during coaching interactions and offline reflective processes that review sessions to assess safety, goal alignment, and coaching quality. Explicit uncertainty estimation guides decision making within a human-in-the loop framework that enables carefully calibrated, progressive autonomy: the system is initially deployed with maximal human oversight, including turn-level monitoring in high-risk or high-uncertainty situations, and reinforcement learning from human feedback is used to incrementally expand autonomy without compromising performance or safety. By integrating adaptive reflection mechanisms, uncertainty-aware reasoning, and structured expert involvement, this approach addresses key limitations of existing systems and ensures that automated coaching remains safe, reliable, and ethically grounded as it scales.
We will leverage anonymized human coaching conversations to identify representative interaction patterns and derive diverse scenarios for system testing and simulated interactions. These scenarios will support active learning and controlled data generation through engagement with the proposed multi-agent system. In parallel, human–human coaching sessions will be evaluated using structured checklists for guideline adherence, response quality, and safety, generating labeled data to establish ground truths, train an initial uncertainty prediction model, and benchmark system performance. Together, these steps enable safe, data-driven development, strengthen reflection-agent capabilities, and provide a robust foundation for reinforcement learning.
Iva Bojic (Nanyang Technological University, Singapore) and Mahnoosh Mehrabani (SoundHound AI, USA) will lead a team of researchers brought together from academia and industry with multidisciplinary expertise to design, build, and iteratively refine a prototype, making systematic improvements until acceptable performance, safety, and reliability standards are achieved. The resulting open-source prototype and findings will provide shared tools, benchmarks, and practical insights to help other researchers in the field further advance safe and effective human–AI collaborative health coaching systems.
