Graphing the Everyday: A Neurosymbolic Approach to Eliciting Routines for Just-In-Time Adaptive Interventions

Just-In-Time Adaptive Interventions (JITAIs) aim to deliver health support at the most opportune moments. While these systems often rely on passive sensor data, they frequently struggle to account for a user's subjective receptivity—their actual willingness and ability to engage. This study explores a neurosymbolic approach to bridge this gap by using conversational agents to elicit daily routines directly from users. The researchers developed a pipeline that uses Large Language Models (LLMs) to extract structured schedule data from unstructured verbal narratives, storing this information in a Neo4j knowledge graph to create a persistent, machine-actionable mental model of the user's day.

A primary contribution of this work is the identification of a fundamental friction point termed the "mental-model gap." While LLMs are highly fluent in conversation, they tend to process information linearly. In contrast, human storytelling is inherently hierarchical and non-linear. When the AI attempts to force these complex, fluid narratives into rigid, deterministic data structures, it often results in "entity fragmentation," where events, times, and locations are incorrectly parsed or disconnected. This misalignment creates an "ecological mismatch," where the system's inferred schedule availability ignores the user's fluctuating psychological and physical energy levels.

To resolve these tensions, the authors propose several design heuristics aimed at evolving rigid schedule-trackers into more empathetic, proactive agents. Key strategies include:

• Routine Piggybacking: Leveraging existing, stable habits to anchor new health interventions.
• Adaptive Negotiation: Allowing the agent to dynamically adjust its timing based on user feedback rather than adhering to a fixed schedule.
• Scalable Transparency: Providing users with visual or natural-language summaries of how the AI has interpreted their routine, enabling them to identify and correct errors in the system's internal model.

Ultimately, the study suggests that by combining conversational fluency with structural transparency, developers can create systems that are not only more accurate but also more aligned with the user's lived experience, thereby supporting more effective long-term behavior change.