When a clinician speaks to a patient, the words are only part of the context. A cough that interrupts a sentence, an audible wheeze, a repeated throat clear — these are all data points that humans use to guide the conversation. A wet, productive cough shifts follow-up questions one way; a persistent dry cough may instead elicit follow-up questions about medication use or reflux. These sounds can change the progress of the conversation itself. Hearing a cough prompts the clinician to ask specific questions, questions that might never come up if the patient doesn’t volunteer the symptom. Patients often don’t. They may normalize their cough, not notice it, or not think it’s relevant to the discussion.

Conversational AI systems that process speech alone miss all of this. They transcribe the words in isolation, which means they lose exactly the signals a clinician would respond to. Building cough detection and subtype classification into the conversation restores a layer of clinical listening: the agent can notice a wet cough the patient never mentioned, ask the follow-up questions a clinician would ask, and flag concerning patterns for human review. This isn’t diagnosis, it’s the more basic skill of paying attention to the patient rather than just the transcript, and it’s a prerequisite for AI agents doing safe, clinically meaningful patient conversations.

HealthCUES goes well beyond detecting that a cough happened. It distinguishes coughing from throat clearing, classifies cough subtypes (dry, wet, barking, whooping)  with confidence scores, and measures how long each event lasts. It does this in real time, with an average end-to-end latency of 340 milliseconds, without interrupting the flow of the conversation. To avoid overwhelming clinicians with alerts, the system also weighs conversational context before flagging anything. For example: a cough during a discussion of respiratory symptoms means something different than one in passing.

For healthcare, the implication is direct: AI agents that conduct patient conversations at scale can now listen the way clinicians do — catching the symptom the patient never mentioned and making sure a human sees it. That’s a small skill with a large surface area, and it’s a necessary step toward conversational AI that is genuinely attentive to patients.