ATLAS: Agenda Tracking and Linking through Adaptive Selection for Streaming Clinical Dialogues under Bounded Context

Clinical conversations contain multiple concerns and questions that emerge, recur, and may remain unresolved. We introduce ATLAS (Agenda Tracking and Linking through Adaptive Selection), a streaming framework that updates a structured agenda state from each utterance under bounded context. ATLAS extracts typed events, associates them with dynamic agendas, links open questions to exact answer utterances or abstains, and derives whether each agenda is unaddressed, addressed, or fully addressed. A confidence router accepts high-confidence NLI candidates, rejects low-confidence ones, and sends only uncertain cases to an LLM verifier. On a 140-conversation split held out from verifier training, confidence routing reduces verifier calls by about 83%. Controlled synthetic conversations augment sparse agenda and unanswered-question patterns. Bounded, learned context selection matches or beats retaining far more context, so task-relevant selection outweighs quantity.