TLDR: SePA (Search-enhanced Predictive AI Agent) is a novel LLM health coaching system that integrates personalized machine learning and retrieval-augmented generation. It uses wearable sensor data to predict daily stress, soreness, and injury risk, then leverages a trusted web-retrieval pipeline to provide proactive, evidence-based, and personalized health advice. An expert study demonstrated that SePA’s retrieval-based advice was preferred over non-retrieval baselines, highlighting its potential for next-generation, trustworthy personal health informatics systems.
In an era where wearable sensors constantly collect vast amounts of personal health data, a significant challenge remains: translating this raw information into actionable, personalized, and trustworthy guidance. While devices from Fitbit, Garmin, and Apple provide metrics and trend visualizations, users often struggle to understand what steps to take next to proactively manage their health and mitigate future risks.
Large Language Models (LLMs) have shown great promise in bridging this gap, offering richer, more personalized insights. However, many existing systems primarily focus on retrospective analysis, explaining past data rather than forecasting future wellness states or providing preventive advice.
Addressing this critical need, researchers Melik Ozolcer and Sang Won Bae from Stevens Institute of Technology introduce SePA (Search-enhanced Predictive AI Agent). SePA is a novel LLM health coaching system designed to offer adaptive, evidence-based guidance by integrating personalized machine learning and retrieval-augmented generation. It aims to move beyond simply explaining what happened to proactively advising on what might happen and what to do about it.
How SePA Works: A Proactive Approach to Health
SePA’s architecture is built on three core components that work together to provide personalized and trustworthy health coaching:
- Proactive Health Predictions: A key innovation of SePA is its ability to forecast daily subjective states such as stress, muscle soreness, and injury risk. Using up to 72 hours of wearable and environmental data, including sleep metrics and vitals, SePA predicts these scores for the day ahead. This allows users, particularly athletes, to receive actionable insights before starting their day. The system employs a two-tiered predictive modeling strategy: a generalized model for new users (cold-start scenario) and more accurate personalized neural models that unlock once a user provides sufficient historical data (typically more than 15 days of self-reported labels). This personalized approach significantly outperforms generalized models, achieving higher accuracy for stress, injury risk, and soreness predictions.
- Trusted, Context-Aware Web-Retrieval Pipeline: When a user seeks advice, SePA’s unique web-retrieval pipeline springs into action. It augments the user’s query with personal context, including demographics and current ML risk predictions. This transforms a generic question into a privacy-preserving, highly specific search prompt (e.g., “Strategies to reduce soreness for a 21-year-old basketball player with high soreness (74%) and elevated RHR”). The system then searches a curated whitelist of 35 trusted domains, including professional societies and major medical centers, ensuring that all retrieved information is from expert-vetted sources. The relevant snippets are then used to ground the LLM’s advice, ensuring it is contextually relevant, reliable, and verifiable.
- Privacy and Transparency: SePA prioritizes user privacy by deleting raw health data after processing and anonymizing all information sent to external APIs. It also offers the option to use no-retention LLM endpoints. The web-retrieval pipeline’s implementation, domain whitelist, and prompt templates are publicly available, fostering transparency and reproducibility.
Validating SePA’s Effectiveness
The researchers conducted a two-part evaluation to validate SePA’s core components. First, they confirmed that personalized predictive models significantly outperform generalized baselines for stress, soreness, and injury risk, especially as more user data becomes available. This underscores the importance of individualization in health coaching.
Second, a blind evaluation with four domain experts (including a psychology professor, head coaches, and a strength and conditioning coach) assessed the quality of SePA’s coaching. Experts were presented with responses generated by SePA with and without the web-retrieval pipeline. The results showed a strong preference for the web-retrieval augmented system (SePA-web), which achieved a superior mean rank and received more first-place votes. This indicates that grounding advice in external, verifiable knowledge significantly enhances the quality, relevance, and helpfulness of health coaching.
Qualitative feedback from the experts further reinforced these findings, praising the system’s ability to combine personalized data with credible web content and its potential to provide privacy for users asking sensitive questions. They also highlighted the importance of detailed, cited answers for building trust.
Also Read:
- Bridging the Gap: Tailoring Health Simulation Explanations with AI for Every Stakeholder
- Optimizing Large Language Models for Clinical Data Extraction
The Future of Personalized Digital Health
SePA represents a significant step towards the next generation of digital health agents. By integrating proactive risk prediction with trusted, context-aware retrieval, it offers a practical blueprint for systems that are not only intelligent but also predictive, transparent, and trustworthy. While acknowledging limitations such as the current cohort of student-athletes and the small expert panel, the research lays a strong foundation for future work in enhancing conversational capabilities and validating the system on larger, more diverse populations.
For more in-depth information, you can read the full research paper: SePA: A Search-enhanced Predictive Agent for Personalized Health Coaching.
