LongCAF: A longitudinal context-constructive agentic framework for LLM-supported hypertension self-management

Long-term hypertension self-management requires patients to make daily decisions and be cautious of when to seek medical attention. Although large language model (LLM)-based systems provide new opportunities for conversational self-management support, conventional question-answering conversational paradigms may be insufficient for providing accurate and personalized responses under conditions of fragmented patient expression, incomplete contextual reporting, and changing interaction intents. Therefore, this study aims to address the current limitations by developing and evaluating a context-constructive agentic framework that integrates longitudinal memory, interaction coordination, proactive communication, and safety escalation.

This study proposes the Longitudinal Context-constructive Agentic Framework (LongCAF) that structures longitudinal patient-LLM interaction as a unified workflow with five core modules: longitudinal context memory for constructing and maintaining patient-specific context, intent-aware interaction coordination for categorizing and routing heterogeneous interaction intents, interaction synthesis for organizing and generating context-grounded personalized responses, proactive communication for eliciting missing information and clarifying patient concerns when possible, and safety escalation for governing and escalating safety-critical interactions to clinicians. The framework was evaluated in a 6-month real-world deployment.

LongCAF supported longitudinal context construction and context-grounded patient-system interaction during deployment. Patient-level longitudinal context completeness increased from 42.9% (IQR 14.3%-57.1%) at baseline to 57.1% (IQR 57.1%-71.4%) at follow-up, in which 79,290 clinically relevant entries were added, and non-blood-pressure domains accounted for 49.3% of newly acquired entries. Patient-facing outputs supported by retrieved patient-specific context accounted for 69% of outputs, with retrieved domains spanning multiple domains. Candidate-question adoption reached 84.4% among eligible interactions, and intent-aware interaction coordination achieved a micro-F1 score of 0.911. End-to-end safety escalation achieved a precision of 0.915, recall of 0.977, and F1 score of 0.945. Clinician and patient evaluations suggested that LongCAF-supported responses were clinically relevant, contextually appropriate, and useful for continued self-management.

The real-world deployment demonstrates the feasibility of LongCAF, highlighting its potential as a practical framework for developing more personalized and clinically governed conversational systems for chronic disease self-management.