A prospective clinical feasibility study of a conversational diagnostic AI in an ambulatory primary care clinic

This prospective feasibility study demonstrates that a conversational AI system (AMIE) can safely and effectively conduct clinical history-taking and generate diagnostic suggestions in a real-world urgent care setting, achieving high patient satisfaction and diagnostic accuracy comparable to primary care providers while requiring no real-time human intervention.

The Gist

Imagine you have a headache, a strange rash, or just a general feeling of being unwell. In the past, you'd call your doctor's office, maybe wait on hold, and then sit in a waiting room for 45 minutes before finally seeing a doctor for a 15-minute appointment.

This paper is about a new experiment where a super-smart AI chatbot (named AMIE) tried to act like a "medical detective" before you even saw the real doctor. The goal was to see if this AI could safely talk to patients, figure out what might be wrong, and help the real doctor be better prepared when you finally meet.

Here is the story of what happened, explained simply:

1. The Setup: A "Medical Rehearsal"

Think of the clinic as a busy airport. Usually, passengers (patients) rush through security and straight to the gate (the doctor's office), often forgetting to pack the right documents.

In this study, researchers set up a rehearsal room before the flight.

• The AI (AMIE): Imagine a very knowledgeable, patient, and polite medical student who never gets tired. This AI talked to 100 real patients via text chat on their computers (not phones) up to 5 days before their real appointment.
• The Mission: The AI asked questions like, "When did the pain start?" "Does it hurt when you move?" and "Do you have a fever?" It built a detailed story of the patient's symptoms.
• The Safety Net: Crucially, a real human doctor was watching the chat from behind the scenes (like a safety supervisor in a control tower). If the AI said anything dangerous or the patient got upset, the human could hit the "stop" button immediately.

2. The Results: Did the AI Pass the Test?

The Safety Score: 10/10
The most important question was: Did the AI hurt anyone or say something crazy?

• Answer: No. The human supervisors watched all 100 conversations and never had to stop the chat. The AI was calm, didn't panic the patients, and didn't suggest dangerous things. It was like a very careful librarian who knows exactly which books to recommend.

The "Detective" Skills: Pretty Good!
After the chat, the AI made a list of possible diagnoses (a "differential diagnosis").

• The Result: In 90% of cases, the AI's list included the actual final diagnosis that the real doctor eventually confirmed 8 weeks later.
• The Comparison: When experts compared the AI's list of possibilities to the real doctor's list, they were equally good. The AI was just as smart at guessing what was wrong as the human doctor.

The "Action Plan": The AI Needs a Little Help
The AI also suggested what to do next (tests, treatments, etc.).

• The Result: The AI's plans were safe and appropriate. However, the human doctors' plans were rated as more practical and cheaper.
• Why? Think of the AI as a brilliant theorist who knows all the rules of physics, but the human doctor is the engineer who knows how to build the bridge with the materials actually available in the local hardware store. The AI sometimes suggested perfect but expensive or complicated solutions, while the doctor knew what was easiest and most affordable for that specific patient.

3. How Did People Feel?

The Patients:

• Before the chat: Some were skeptical. "Can a robot really understand my pain?"
• After the chat: They felt much better. They said the AI was polite, listened well, and didn't make them feel rushed. It was like having a conversation with a friend who really cares.
• Trust: Their trust in AI went up significantly. They felt more prepared for their real doctor visit.

The Real Doctors:

• When the patients arrived for their real appointment, the doctors had already read the AI's chat summary.
• The Feeling: Doctors said, "Wow, this is helpful." Instead of spending the first 10 minutes just gathering basic facts, they could jump straight into the important stuff. It was like arriving at a meeting where everyone had already read the agenda beforehand.
• The Vibe: The visits became more like a partnership rather than an interrogation.

4. The Catch (The Limitations)

While the results are exciting, there are a few "buts":

• The Tech Barrier: The study required patients to use a laptop or desktop computer. People who only had smartphones or weren't good with computers were left out. This is like testing a new car only on smooth highways and not on bumpy dirt roads.
• No Physical Touch: The AI couldn't feel a lump, listen to a heartbeat, or look in an ear. It's a great detective, but it can't do a physical exam.
• The "Hawthorne Effect": Because patients knew a human was watching the AI, they might have been extra careful with their answers. In the real world, without a human watching, the AI might act slightly differently.

The Big Picture

This study is like a successful test drive for a self-driving car.

• It proved that the car (the AI) can drive safely on the road (the clinic) without crashing.
• It proved that the passengers (patients) feel comfortable in the car.
• It proved that the car can help the human driver (the doctor) get to the destination faster.

The Conclusion:
We aren't ready to replace doctors with robots yet. But this study shows that AI can be a super-powerful assistant. It can do the "homework" of listening and organizing information so that when you see your doctor, you can spend that precious time talking about what matters most, rather than just filling out forms.

It's a huge step toward a future where your medical care is faster, safer, and less stressful for everyone.

Technical Summary

Here is a detailed technical summary of the paper "A prospective clinical feasibility study of a conversational diagnostic AI in an ambulatory primary care clinic."

1. Problem Statement

The primary care sector faces a global shortage of physicians, leading to increased workloads, burnout, and reduced access to care. While Large Language Models (LLMs) have shown promise in simulated clinical environments (e.g., Objective Structured Clinical Examinations with actors), their translation into real-world clinical workflows remains unproven. Key challenges include:

• Safety: Ensuring AI does not cause harm, undertriage critical conditions, or provide unreliable advice, particularly in unsupervised settings.
• Workflow Integration: Determining how AI can effectively integrate into existing Electronic Health Record (EHR) systems and provider workflows without disrupting care.
• Clinical Reasoning: Verifying if AI can perform diagnostic history-taking and generate differential diagnoses (DDx) and management plans (Mx) comparable to human Primary Care Providers (PCPs) in complex, unstructured real-world scenarios.
• Trust: Assessing patient and clinician acceptance of AI as a diagnostic partner.

2. Methodology

The authors conducted a prospective, single-arm feasibility study at Healthcare Associates (HCA), an academic primary care practice within Beth Israel Deaconess Medical Center (BIDMC) in Boston, MA.

• System (AMIE): The study utilized the Articulate Medical Intelligence Explorer (AMIE), an LLM-based conversational agent built on the Gemini 2.5 family of models (specifically 2.5 Pro and Flash) with "Thinking Mode" enabled. The system was not fine-tuned on domain-specific data but was aligned via prompt engineering and state-aware reasoning strategies.
  • Workflow: AMIE operates in five phases: Intake, History Taking, Diagnostic Validation, Deliver Assessment, and Wrap-up. It maintains an internal state (patient summary, working DDx, information gaps) to guide the conversation.
• Study Design:
  • Participants: 100 adult patients (established HCA patients, age 18+, English-speaking) scheduled for urgent care visits.
  • Process:
    1. Pre-Visit: Patients engaged in a synchronous text-chat with AMIE 0–5 days before their appointment.
    2. Safety Oversight: A board-certified internal medicine physician (AI Supervisor) monitored all interactions in real-time via screen-sharing. Pre-defined safety criteria (e.g., self-harm risk, clinical harm potential) were in place, though no interventions were triggered.
    3. Provider Handoff: AMIE generated a summary and chat transcript (including a list of potential diagnoses) which was provided to the PCP prior to the visit. Crucially, the AI's management plan was hidden from the patient and PCP for research evaluation only.
    4. Post-Visit: 8 weeks after the visit, a retrospective chart review established the "ground truth" final diagnosis.
• Evaluation Metrics:
  • Safety: Number of safety stops/interventions.
  • Diagnostic Accuracy: AMIE's DDx compared to the ground truth final diagnosis (Top-k accuracy, Bond/Graber scale).
  • Quality Assessment: Blinded, randomized comparison of AMIE's and PCP's DDx and Mx plans by a panel of 8 board-certified internal medicine physicians.
  • User Experience: Surveys (GAAIS, GMCPQ, PACES) and semi-structured interviews with patients and providers.

3. Key Contributions

• First Real-World Prospective Study: This is the first study to evaluate a generative AI diagnostic agent in a live, high-volume urgent care setting with real patients, moving beyond simulated actors.
• Safety Framework: Implementation of a rigorous, real-time human-in-the-loop safety oversight protocol for a system that directly presents potential diagnoses to patients.
• Blinded Comparative Evaluation: A robust methodology for comparing AI and human performance, including automated formatting and truncation of outputs to ensure blinded, unbiased rating by clinicians.
• Diagnostic Reasoning Analysis: Detailed analysis of the AI's internal reasoning traces (confidence, entropy, differential evolution) over the course of the conversation.

4. Key Results

A. Safety and Feasibility

• Zero Safety Stops: Across 100 patient-AMIE interactions, no safety interventions were required by the human supervisors.
• Completion Rate: 87.7% of consented patients completed the AMIE interaction (100/114). Failures were primarily due to technical issues (device compatibility) or protocol deviations (e.g., pregnancy discovered during chat).
• Trust: Patient attitudes toward AI improved significantly after interacting with AMIE (p < 0.001), with increased perceived utility and reduced concerns.

B. Clinical Reasoning Performance

• Diagnostic Accuracy:
  • AMIE included the final ground-truth diagnosis in its top-7 candidates in 90% of cases.
  • Top-3 Accuracy: 75%.
  • Top-1 Accuracy: 56%.
  • Performance remained high even when the final diagnosis was confirmed by diagnostic testing (N=46).
• Quality of DDx and Management Plans (Blinded Assessment):
  • Overall Quality: Clinical evaluators found no statistically significant difference between AMIE and PCPs in the overall quality of Differential Diagnoses (p=0.6) or the appropriateness and safety of Management Plans (p=0.1 and p=1.0, respectively).
  • Areas of AI Inferiority: PCPs significantly outperformed AMIE in the practicality (p=0.003) and cost-effectiveness (p=0.004) of management plans. This is attributed to AMIE lacking access to the patient's EHR, physical exam findings, and real-world resource constraints.
  • Blinding: Evaluators could only guess the source (AI vs. Human) correctly 59% of the time, indicating the outputs were sufficiently similar to prevent provenance bias.

C. User Experience

• Patients: Rated AMIE highly for empathy, politeness, and listening skills. They appreciated the detailed history-taking and felt better prepared for their visits. However, trust in data confidentiality and honesty remained a concern.
• Providers: 75% of PCPs found the AMIE summary helpful for visit preparation. They noted that patients arrived more organized, allowing visits to shift from data gathering to collaborative decision-making.

5. Significance and Future Directions

• Clinical Translation: The study demonstrates that conversational AI can safely and effectively perform pre-visit clinical history taking in a real-world setting, serving as a viable "digital scribe" or intake tool.
• Human-AI Collaboration: The results suggest a future where AI handles the initial, time-consuming data gathering and preliminary reasoning, allowing human providers to focus on verification, complex decision-making, and the therapeutic relationship.
• Limitations & Next Steps:
  • Context Gap: AMIE's lower scores in practicality/cost-effectiveness highlight the need for multimodal integration (EHR access, physical exam data) to improve clinical utility.
  • Safety: While safe under strict human supervision, the study did not test fully autonomous deployment. Future work must define scenarios where AI can operate with "human-on-the-loop" rather than "human-in-the-loop."
  • Equity: Technical barriers (device access, digital literacy) excluded some demographics, pointing to a need for better onboarding and accessibility features.

In conclusion, this study provides the first empirical evidence that a state-of-the-art LLM can conduct safe, high-quality diagnostic conversations with real patients, achieving diagnostic accuracy comparable to human physicians, while identifying specific areas (practicality, cost) where human oversight remains superior.