A Modular Reference Architecture for Patient-Facing Generative AI Health Coaches in Chronic Self-Management

Patient-facing generative AI is increasingly proposed for chronic self-management support, but trust-relevant properties such as source verification, safety boundaries, traceability, bias monitoring, and escalation are difficult to inspect when treated as emergent properties of a dialogue model. This paper presents LUCID, a modular conceptual reference architecture for non-diagnostic, patient-facing generative AI health coaching grounded in Therapeutic Patient Education (TPE). It separates curated knowledge, controlled health-context integration, analysis and personalization, response generation through an internal large language model (LLM) Gateway, safety, explainability, audit, and staged evaluation. The architecture mediates patient requests, preferences, symptom reports, wearable summaries, educational goals, source-linked explanations, and escalation notices through explicit responsibility boundaries. Six design requirements are derived from a related-work synthesis, and key decisions are documented through Architecture Decision Records (ADRs). A runtime walkthrough and staged evaluation plan show how the architecture can be tested from architecture readiness to human-centered feasibility. By making curated sources, released context fragments, competency targets, model and prompt versions, safety and release decisions, and escalation paths explicit, LUCID enables developers and evaluators to inspect and compare how individual patient-facing responses are produced.