An AI Intake Agent Captured More Complete Eye-Care Histories in a Randomized Trial

Medical history-taking is where a patient’s lived experience becomes usable clinical evidence. The blur that began at night, the old surgery that suddenly matters, the treatment that failed months ago—those details can shape what a physician looks for next.

A randomized clinical trial at Zhongshan Ophthalmic Center in Guangzhou, China, tested whether an AI intake agent could preserve more of that story before a non-emergency eye appointment. Among 172 patients, histories gathered by the agent scored substantially higher than histories gathered by ophthalmology residents on the study’s 100-point MedHistory rubric.

The result is not evidence that an AI system is a better doctor. It is evidence that a carefully engineered system can perform one bounded task—patient interviewing and structured history capture—with unusual thoroughness, leaving the physical examination, diagnostic synthesis and treatment decision with clinicians.

Research references: [1]

A stronger history, built before the exam

Researchers enrolled patients between May 10 and June 10, 2025. Of 175 people screened, 172 were randomized equally: 86 completed pre-consultation with the AI agent and 86 with an ophthalmology resident. There were no withdrawals after randomization, and the paper reports no adverse events.

The primary outcome was the MedHistory composite score, evaluated by two senior ophthalmologists who were masked to group assignment. The agent group averaged 80.73 out of 100, compared with 56.94 for the resident group. After adjustment for age, sex and pre-consultation duration, the mean difference was 17.67 points (95% confidence interval 14.07–21.27; P<0.001).

That adjustment matters because the AI conversations were much longer. Median pre-consultation time was 11.2 minutes with the agent and 3.1 minutes with a resident. The study therefore does not show that AI made intake faster. Its promise is different: a system can provide more time for systematic questioning without consuming the same block of resident time.

The MedHistory framework was developed for this study through ophthalmologist focus groups and prior process-oriented work. Its 100 points heavily weighted history-taking, with smaller components for communication and documentation. That makes it useful for the task the researchers designed—but it is not yet a widely validated clinical-outcomes instrument.

Research references: [1] [2]

Patients experienced patience, not speed

Participants rated the agent higher for perceived patience and empathy. On five-point scales, the median patience rating was 5 for the agent and 4 for residents; perceived empathy was 5 versus 3. Overall satisfaction was high and did not significantly differ between groups: a median of 5 in both.

Those ratings support careful language. “Patients felt heard” is an appealing interpretation, but it was not a named trial endpoint. The measured findings were higher ratings for patience and perceived empathy after a longer interaction.

The human picture was also mixed in an important way. Residents were rated more effective at reducing anxiety: a median 4.5 compared with 3 for the agent. The trial therefore suggests that an AI system can feel patient and methodical without reproducing every dimension of human reassurance.

Research references: [1]

The system was designed for one job

This was not a general chatbot placed in a clinic. The researchers built a locally deployed system based on a fine-tuned Qwen-2.5-72B model. A main agent conducted an adaptive interview using symptom-oriented logic trees. An auxiliary agent extracted clinical details and generated a preliminary report, while a documentation module standardized terminology and formatting.

For a complaint such as red eye, the questioning structure could move through vision changes, pain, discharge and trauma. The design attempted to separate comprehensive information gathering from premature diagnostic closure.

That architecture helps explain both the result and its limits. The trial tested the whole system—model, fine-tuning, prompts, logic trees, memory, interface, longer interview and documentation pipeline. It cannot isolate how much of the advantage came from any one component.

The documentation subscore itself did not significantly differ from residents after adjustment. The agent scored higher on communication and especially on detailed history capture, while residents used medical terminology more accurately in item-level analysis. A later optimization step reduced speech-to-text and wording artifacts in the study system.

Research references: [1] [2]

Physical findings changed the comparison

Exploratory analyses offer a useful picture of human–AI division of labor. Based on history alone, the agent’s primary-diagnosis F1 score was 0.85, compared with 0.68 for residents. But when ocular examination signs were added, residents improved more, and the diagnostic gap narrowed. Test-recommendation performance did not significantly differ between groups across the assessed stages.

Treatment-plan quality after ocular signs were integrated also showed no detected difference: both groups had a median rating of 3 on a three-point scale. This was not an equivalence trial, so “no significant difference” should not be read as proof that the approaches are interchangeable.

The reference standard for these exploratory comparisons was based on the initial specialist workup rather than final longitudinal diagnostic confirmation. The paper itself cautions against treating the diagnostic analysis as evidence of clinical superiority.

The more credible near-term workflow is complementary: AI gathers and structures what the patient reports; the clinician examines the eye, interprets physical findings, interrogates the prepared record and remains responsible for care.

Research references: [1] [2]

What this trial established—and what it did not

The study moved beyond simulated cases. Eighty-six real patients completed AI-led pre-consultation in a live specialty clinic, and masked specialists evaluated the resulting histories. The trial was prospectively registered as NCT06824389, and the paper reports that its protocol and statistical analysis plan are included in the supplementary material.

But the evidence comes from one tertiary hospital, one specialty, one month of enrollment and a Chinese-language, non-emergency setting. The trial was open-label. Its longer AI interviews may have contributed materially to the higher history scores, and the design cannot distinguish novelty effects from lasting patient preferences.

The study did not measure whether physicians saved time, experienced less cognitive load, avoided more errors or achieved better patient outcomes. It did not test emergency intake, high-volume deployment, other languages or patients whose communication needs differ from those represented here.

The paper reports no adverse events, but 172 participants cannot establish safety for broad clinical deployment. Any implementation would still need human review, audit trails, clear escalation pathways, privacy protections and monitoring for systematic omissions or misleading summaries.

Patient-level data are not public because of privacy and institutional restrictions. The authors made core system and analysis code available on Zenodo, while withholding institutional integration code for security and privacy reasons. The study lists public and institutional funding and declares no competing interests.

Research references: [1] [3]

More of the patient can arrive in the room

The most meaningful possibility here is not an autonomous eye doctor. It is a better handoff into the human encounter.

A patient could have space before the appointment to remember when symptoms began, describe what changed and clarify what has already been tried. A clinician could begin with a structured account that is reviewable rather than a thin form or a hurried fragment.

That promise still has to survive pragmatic testing. Future studies should measure the entire visit, not only pre-consultation: duplication, total encounter time, clinician workload, missed details, error patterns, patient trust and health outcomes. They should also test whether the advantage persists when the comparison group has equivalent interview time.

For now, the Guangzhou trial offers a concrete advance. It shows that an AI intake agent can gather more complete eye-care histories from real patients on the study’s rubric, while the later examination and clinical judgment remain human work. The machine’s role is not to replace the encounter. It is to help more of the patient’s story reach it.