An AI conversation before a specialist appointment helped patients present their histories and made the subsequent consultation shorter in a randomised trial at two Chinese health centres. The preassessment tool, PreA, collected information and generated a referral report for the clinician.[1]
The Nature Medicine trial randomised 2,138 participants, with 2,069 in the final analysis: 1,620 patients and 449 care partners. It covered 24 disciplines and 111 specialists. In the independent-use group, mean consultation duration was 3.14 minutes, compared with 4.41 minutes without PreA—a reported 28.7% reduction. Participants also gave higher communication ratings, making the result about the quality of the handoff as well as its speed.[1][2]
Better conversation, measured rather than imagined
The authors measured more than speed. Patients using PreA independently rated ease of communication at 3.99 on a five-point scale, compared with 3.44 for people receiving usual care—a relative difference reported as 16%. Specialists rated the value of the referral for coordination at 3.69 rather than 1.73. The study also reported improvements in patient satisfaction, perceived attentiveness and interpersonal regard. Those ratings were self-reported by participants who knew which group they were in, a reasonable caution when interpreting an experience rather than a laboratory endpoint.[1][2]
A chatbot taking a history sounds like a trivial replacement for a paper form until one asks what the form cannot do. A conversational interface can invite a patient to describe a symptom in ordinary language, seek missing context and arrange that information for someone who has only minutes to read it. PreA was developed with community participants, medical staff and other local stakeholders, including people who needed low-literacy support. It was designed for a specific clinical handoff, not to roam through medicine acting as an autonomous doctor.[1]
Participants were randomised to PreA alone, PreA with staff support, or no PreA. The two PreA groups performed similarly on several measured outcomes. Independent use matters operationally because a tool that needs a staff member for every interview moves work rather than necessarily releasing staff time.[1]
The clock has two hands
Patients spent about 3.5 minutes using PreA beforehand, while the specialist consultation was 1.27 minutes shorter. The result therefore shows a redistribution of time and information within the visit, rather than a shorter total patient journey. Waiting times, roughly 34 minutes in both comparisons, were not significantly different. A separate comparison found that participating physicians saw more patients per shift than matched nonparticipants, but that throughput analysis was not the main participant-level randomisation.[1][2]
Specialists remained responsible for diagnosis and treatment. The trial assessed communication, coordination and consultation time; longer-term clinical outcomes were outside those endpoints.[1]
A small design decision with large implications
The researchers also ran a separate simulation comparing their community-co-designed system with a version further fine-tuned on collected local clinical conversations. Unexpectedly, additional local training made the model worse in several graded domains. The team argued that imitating existing practice can reproduce its weaknesses. The lesson is not that local data are bad, but that the goal of a healthcare tool is to help people, not merely mimic what commonly happens in a pressured clinic.[1]
A clinician starting with a more complete account can spend the appointment differently. The useful advance here was a better-prepared conversation, delivered through a specific pre-consultation workflow.[1][2]
Sources & further reading
- An LLM chatbot to facilitate primary-to-specialist care transitions: a randomized controlled trial Nature Medicine · 19 Jan 2026 · Accessed 24 Oct 2026
- An LLM chatbot to facilitate primary-to-specialist care transitions PubMed · Accessed 24 Oct 2026