In a fictional 10-minute consultation, a clinician hears "cough", thinks "viral upper respiratory tract infection" and stops considering alternatives. That illustrates premature closure: an early explanation ends the search before the relevant uncertainty has been resolved. A differential-diagnosis tool may help reopen the question, but a longer list is not a diagnosis or a complete care plan.
This guide was reviewed on 10 October 2026. AMIE's real-patient feasibility publication changes the research picture, while current product descriptions also require a more precise distinction between diagnostic support, clinical reference and regulated functions. The comparison is published by iatroX and includes its own Brainstorm, reference and learning tools. No head-to-head performance test of these products is reported here.
Why DDx tools are attracting attention again
The attraction is understandable: a tool can prompt a clinician to consider an explanation that was not initially prominent. Availability bias, illustrated by favouring the diagnosis seen yesterday, and premature closure are useful concepts for examining how an initial interpretation can become too narrow. The clinical-reasoning discussion in the March 2026 AMIE report places these issues within the wider diagnostic literature.
The original short-consultation example also highlights workflow pressure and multimorbidity. It is an illustration, not a claim that every UK consultation has the same duration or that a model automatically compensates for workload.
A differential generator can be used as a cognitive forcing aid: an invitation to reconsider the case. It should not become a machine for adding rare "zebra" diagnoses regardless of context. The useful question is what additional information would make an alternative more or less plausible and whether considering it changes the next decision.
The four categories of DDx support
Classic differential generators: checklist expansion. Isabel DDx Companion is a relevant example. Its product description, checked on 10 October 2026, supports entering presenting features, receiving suggested conditions and identifying important conditions not to miss. A fictional input such as "abdominal pain, travel, fever" illustrates the starting point, not a sufficient history for a diagnosis. Isabel's current product page also describes linked clinical knowledge, so it should not be reduced to a bare list generator.
The previously cited approximately 95% figure needs a specific attribution. Isabel's accuracy-study overview, checked on 10 October 2026, says that in general the studies it summarises found the final diagnosis in its list in 95% of cases from initial clinical features. That is a vendor summary of list inclusion, not a current first-choice accuracy estimate for every clinical setting. It cannot be ranked directly against AMIE's differently designed diagnostic analysis.
Visual and pattern tools: comparing appearances. VisualDx's October 2026 public description combines a differential builder with clinical images, look-alikes and structured condition information. That can address a different gap from a text-only explanation, particularly when a clinician needs to compare morphology rather than remember a textbook phrase.
The fictional "annular rash on the shin" example invites comparison of ringworm and granuloma annulare. Erythema nodosum can also be useful in teaching lower-leg appearances, but as a contrasting nodular presentation, not an interchangeable annular-rash label. DermNet's granuloma annulare reference and erythema nodosum reference, checked on 10 October 2026, help preserve that distinction. Image comparison supports assessment; it does not establish a diagnosis without the relevant clinical information.
The earlier guide's "gold standard" wording for dermatology and ophthalmology should not be read as a universal comparative result. VisualDx may be the more suitable resource when the task specifically requires a clinical image library, but this guide has not established superiority across all skin, eye or general diagnostic tasks.
LLM-based diagnostic assistants: exploring a vignette. DxGPT's current website, checked on 10 October 2026, describes support for generating diagnostic hypotheses, including complex and rare-disease contexts. A fictional paediatric or genetics-oriented vignette may therefore be a different use case from checking a straightforward management recommendation.
The illustrative phrase "consider Kawasaki disease because of five days of fever and a rash" is a prompt to examine a possibility, not a complete diagnostic rule and not a measured DxGPT response. The relevant clinical features still need to be established and assessed. A plausible explanation cannot substitute for that work.
Conversational diagnostic research systems: active history taking. AMIE investigates a further task: asking questions and refining an account through dialogue. Its earlier simulated research is no longer the whole evidence base. The Lancet publication of 8 October 2026 reports supervised pre-visit conversations with real primary-care patients, followed by their normal appointments. That is research evidence about a defined pathway, not confirmation of open, unrestricted NHS clinical availability.
What the October AMIE study adds
The published abstract reports 114 enrolled patients, with 98 completing both the AI interaction and primary-care appointment. It records no predefined conversation safety stops, one hallucination noted by supervisors and added clinical information in five interactions. The physician safety supervisor and the treating clinician were both part of the evaluated pathway. The October 2026 abstract should be read with those human contributions visible.
The detailed March 2026 diagnostic analysis reported approximately 56% top-one, 75% top-three and 90% top-seven performance. Its matching threshold included exact or sufficiently close diagnoses, and the eventual diagnosis came from follow-up chart review. The preprint is the source for those detailed methods where the final full manuscript could not be inspected directly.
This is an advance beyond actors in controlled consultations. It is not evidence that an AI independently made the right first diagnosis in 90% of cases, that patients received autonomous treatment or that total GP workload fell. iatroX's dedicated AMIE Lancet explainer and diagnostic-statistic article examine those questions in more detail.
Is DxGPT regulated? The version matters
The January 2026 version of this guide described DxGPT as decision support rather than a medical device. That categorical wording is no longer supported by the current vendor description.
As checked on 10 October 2026, DxGPT's FAQ says its professional version has been considered medical-device software for diagnostic support and classified as Class IIa under the relevant EU rule, based on information submitted to AEMPS. It also says the applicable conformity assessment and regulatory process must be completed. This is the vendor's stated position, not independent confirmation of completed certification or permission for every proposed UK use.
The appropriate question is therefore which version, intended purpose and deployment are being considered. "Decision support" and "medical device" should not be treated as mutually exclusive labels. Professional judgement and organisational approval remain separate considerations from the terminology used on a product page.
The safety reality: what the literature supports
Patient-facing symptom checkers and clinician-facing differential tools should not be conflated. A systematic review published in npj Digital Medicine on 17 August 2022 found variable and generally low diagnostic and triage accuracy among the evaluated symptom checkers. It raised concerns about both inappropriate escalation and missed urgency. That historical review does not establish the performance of every product available in October 2026, and "triaging everyone to A&E" would be an inaccurate generalisation.
The original guide also referred to modest gains from diagnostic support and favoured using it early. A multicentre Isabel experiment reported in BMJ Quality & Safety in 2022 found improvements in correct-diagnosis inclusion with both early and late use. Early use broadened hypotheses more, but the study does not justify a blanket claim that early use is always superior for every diagnostic outcome.
The practical interpretation is to use support for a defined purpose, whether broadening an initial differential or challenging a later conclusion. Do not use a tool merely to obtain agreement with an answer already preferred. Nor should any workflow in this article be understood as a guarantee of safety or protection from professional consequences.
A seven-step UK clinician workflow
1. Define the problem representation. Write a concise account before asking for suggestions. The original fictional example, a 45-year-old woman with acute right upper quadrant pain and fever after a fatty meal, is a starting representation, not a complete assessment. Preserve features that do not fit the initial explanation.
2. Use a differential expander for a specific gap. Isabel or iatroX Brainstorm can be considered for broadening a fictional or appropriately governed case discussion. The purpose is to identify alternatives worth assessing, not to outsource the conclusion. Product suitability and information-governance approval must be established separately.
3. Generate discriminating questions. An educational prompt such as "Which three questions would best distinguish the leading two possibilities?" turns a list into a reasoning task. Check whether the proposed questions actually change the assessment and whether the patient has already answered them.
4. Consider relevant conditions not to miss. Aortic dissection and meningitis remain useful examples of serious diagnoses whose relevance depends on the presentation. They are not mandatory additions to every differential. Ask what supports concern and what information is still needed rather than treating a red-flag label as a completed assessment.
5. Decide what investigation or referral question needs answering. Asking for a "gold standard test" can be educational, but the definitive test is not automatically the appropriate first step in primary care. The next action depends on the clinical situation, available pathway and the information already established.
6. Verify decision-critical claims. Check the underlying NICE, CKS or other appropriate UK guidance and current medicines information, including the relevant population and exceptions. A real citation may support only part of a generated recommendation. The GMC's guidance on information sufficiency and consultation mode, checked on 10 October 2026, provides the professional context rather than an endorsement of any tool.
7. Document what was actually considered and done. A fictional record structure such as "Differential considered: X, Y and Z; safety-netting discussed for the relevant concern" is only appropriate when that work occurred. Do not generate a polished record of questions, findings or discussions that never happened.
For public educational tools, use fictional or appropriately de-identified material. Removing a name or NHS number alone may not remove identifiability. Actual patient information belongs only in systems and workflows approved for that purpose; a familiar brand or convincing privacy statement is not sufficient permission.
Where iatroX fits
The original guide described Brainstorm's intended diagnostic arc: targeted history questions, a differential considered in UK context, investigation questions relevant to the NHS and links towards management guidance. That remains a useful way to describe the learning task, not proof that every generated rank or suggested investigation is appropriate.
As described in October 2026, Ask-iatroX provides free clinical reference grounded in NICE, CKS, SIGN and SmPC information from emc, with linked sources. The published methodology describes retrieval, ranking, citation grounding, output checking and uncertainty handling. These are design features, not a universal accuracy guarantee or an independently demonstrated advantage over every competitor.
The practical value is the connection between expanding a question, inspecting its evidence and practising the reasoning it requires. It is not that other tools only generate lists: Isabel and VisualDx also provide clinical context and supporting information, while conversational systems can ask follow-up questions.
Verdicts for different clinical tasks
For a clinician who wants an established differential checklist and already has access, Isabel is a sensible resource to examine. For a task requiring comparison with clinical images, VisualDx offers a form of support that a text discussion cannot replace. For exploratory complex-case hypotheses, DxGPT may be relevant within the limits of its exact version and intended use.
For UK source checking and structured learning, iatroX offers a different combination of functions. For someone assessing the future of conversational intake, AMIE supplies important research rather than a directly interchangeable everyday NHS product. None of these choices eliminates the need to assess the patient, verify consequential claims and decide what happens next.
Turn this clinical topic into active learning
Choose a fictional case, record the initial differential, identify the question that changed it and check the relevant guidance. Then explain what would make the preferred interpretation wrong. That produces a more useful learning record than saving a long AI-generated list.
As described in October 2026, iatroX's CPD workflow connects focused learning with a draft reflection and suggested framework mapping that the learner reviews and personalises. It supports PDF export and direct FourteenFish export for linked accounts, with completed evidence remaining accessible after the subscription ends. CPD tools are included in the paid subscription, not a separate add-on; a learning record is not a claim of automatically accredited CME or proof of clinical competence.
Frequently asked questions
Which AI tool is best for differential diagnosis for doctors?
There is no universal winner: checklist expansion, visual comparison, exploratory reasoning and UK source checking are different tasks. Choose according to the required function, existing access, evidence and approved clinical workflow.
Does Isabel improve diagnostic accuracy?
Published studies have found improvements in the likelihood that the correct diagnosis appears in a differential under their tested conditions. That is not a guarantee of first-choice accuracy, improved outcomes or benefit in every deployment.
How should GPs use differential-diagnosis tools safely in the UK?
Use them to broaden or challenge reasoning, verify decision-critical guidance and retain responsibility for assessment and the plan. Protect patient information and do not treat an AI-generated list or this proposed workflow as a safety guarantee.