The questions that shape a dementia workup are rarely the dramatic ones. Does he shout and thrash in his sleep. Has she stopped paying the bills, or just stopped enjoying it. Was the fall on the stairs a trip, or did the legs simply stop cooperating. Each answer narrows the field, and each one takes minutes that a clinic visit does not have. So a team from Emory University and the University of California, San Francisco tried something blunt: put the questions to patients and their informants before the appointment, using a machine that never checks the clock. The result, published in the peer-reviewed proceedings of the 27th Annual Meeting of SIGDIAL in Atlanta this August (pages 62 to 78), is one of the first tests of a voice AI doing dementia history-taking with real patients in a real clinic.
The system is an LLM-based voice agent working through a script written by dementia specialists, with conditional branching so that a mention of hallucinations or a fall opens the follow-up questions a neurologist would ask next. It was built for the people using it: room to finish a sentence, no interruptions, supportive cues and clarifications when a question does not land. In the study, 30 participants from a cognitive neurology clinic sat with the agent, patients and informants such as a spouse, at the time of their neuropsychological testing. The specialist consultation came the same day or up to several months later, and the clinician was kept blinded to what the agent had been told, conducting the interview exactly as usual.
The headline numbers come from the same team's preprint of the work, posted on arXiv in September 2025 and not itself peer reviewed. Across 25 paired agent and clinician transcripts, human annotators (agreeing with each other at a Cohen's kappa of 0.72 to 0.73 before consensus) scored what each interview had surfaced. The agent reached an overall sensitivity of 83.5 percent, with a 95 percent confidence interval of 76.3 to 88.7, and specificity of 100 percent, meaning that when it recorded a symptom as present it was rarely putting words in anyone's mouth. Its more interesting advantage was coverage: across 21 symptoms the specialists themselves had picked as clinically important, the median rate of a symptom simply never coming up was 13.0 percent for the agent and 30.4 percent for the clinicians. The machine asks everything because it can. A doctor asks what the visit has room for.
The cracks are in the detail, and the authors do not hide them. How the model was prompted changed what it heard: a sequential strategy (11 interviews) reached 87.6 percent sensitivity against 78.7 percent for a full-script version (12 interviews), which is a large gap resting on very small groups. Symptom by symptom, among the 18 items with at least five usable cases, sensitivity ran anywhere from 40 percent to 100 percent, and most of those estimates carried confidence intervals more than 40 points wide. The paper describes itself as exploratory, with no prespecified endpoints, and treats its own results as hypothesis generating. That is the correct label, and it should travel with the numbers wherever they go.
For a family, the practical shape of this is worth understanding, because the appeal is obvious and the limits are less so. What the agent produces is a record of a conversation, raw material a clinician reads before or during the visit; the study's Figure 1 frames it exactly that way, with the transcript going to the clinician to support the diagnostic decision. It does not decide anything. In this study it was trained annotators, not software and not a doctor under time pressure, who pulled the symptom list out of those transcripts against a consensus standard, so the last step of the pitch, a genuinely physician-ready summary that holds up in a busy clinic, is still an open question. Nothing here measured whether a diagnosis came sooner, was more accurate, or changed what happened next for anybody.
Still, the direction is telling. The observations that matter most in dementia care are accumulated at home, over months, by people who notice the second missed appointment and the third bad night, and they are usually compressed into whatever fits between the blood pressure cuff and the door. A patient interview that runs long by design, in a waiting room, with the answers written down, is not a clever piece of engineering so much as an admission about how little room the visit itself has left.