Wearable Physiological Sensing and Retrieval-Augmented Large Language Models for Personalized Mental Health Dialogue: A Proof-of-Concept Investigation of HRV-Grounded Conversational Support
Main article
Abstract
Background: Conventional approaches to AI-driven mental health support rely exclusively on conversational text, leaving a critical personalization gap: the system interacts with the user but has no knowledge of the user's physiological state. Wearable electrocardiography (ECG) and photoplethysmography (PPG) sensors now provide continuous, passively acquired heart rate variability (HRV) data that are well-established proxies for autonomic nervous system regulation and affective state. Translating these physiological signals into clinically interpretable, dialogue-ready knowledge represents an underexplored pathway to genuinely personalized mental health conversational support.
Objective: This study introduces physiological grounding, defined as the semantic interpretation of continuously acquired HRV and autonomic biomarker data to serve as retrievable context for large language model (LLM) dialogue personalization, and evaluates its feasibility through a two-study proof-of-concept investigation.
Methods: Study 1 applied a combined data-driven and theory-driven feature selection procedure to a multi-site wearable ECG and PPG dataset (N = 312 participants, 4,680 recording days) to identify 52 psychologically interpretable HRV and autonomic features organized into a semantic knowledge base. Study 2 conducted an 8-week longitudinal investigation (N = 42 participants) in which a physiologically grounded LLM dialogue system provided personalized mental health conversational support, with pre-post assessments of depression (PHQ-9), anxiety (GAD-7), perceived stress (PSS-10), well-being (WEMWBS), and subjective happiness (SHS).
Results: Ensemble classification achieved F1 scores of 0.68–0.97 across affective state labels and time windows, with the RMSSD and LF/HF ratio features demonstrating the strongest discriminative power. Post-intervention, PHQ-9 scores declined by 2.7 points (p = 0.003), GAD-7 by 2.1 points (p = 0.009), and SHS increased by 3.6 points (p < 0.001). Thematic analysis revealed that physiological awareness produced a qualitatively distinct "body-informed" personalization that users perceived as more trustworthy than text-only dialogue personalization.
Conclusion: Physiological grounding is a semantically and experientially feasible strategy for personalizing mental health LLM dialogue. The interpretive bridge from HRV signals to dialogue-relevant affective knowledge is functional, and users perceive physiologically grounded responses as meaningfully personalized. Randomized controlled evaluation and real-time wearable integration are the critical next steps.
