iatroX: UK AI Clinical Reference Platform

• iatroX is a UK-centred AI clinical reference platform that synthesizes rapid, evidence-grounded answers from trusted sources like NICE and SIGN.
• It employs a safety-aware, retrieval-augmented generation pipeline with explicit provenance to align outputs with accepted UK practices.
• The platform shows high real-world engagement and positive clinical feedback, effectively mitigating information overload in time-pressed environments.

iatroX is a UK-centred AI clinical reference platform designed to provide rapid, evidence-grounded answers to clinical questions at the point of care for UK healthcare professionals. It is presented as a generative AI clinical decision support/reference system rather than an autonomous diagnostic engine, and its core design objective is to mitigate clinical information overload while avoiding the uncited, non-local, and potentially hallucinatory outputs associated with unconstrained general-purpose LLMs. The platform is built around a retrieval-augmented generation workflow grounded in trusted UK clinical sources and is evaluated in an early mixed-methods real-world implementation study covering web, iOS, and Android usage over 16 weeks in 2025 (Tytler, 25 Sep 2025).

1. Definition, scope, and intended role

iatroX is intended for clinicians working in UK healthcare settings who require rapid access to guidance during live clinical workflow. Its stated function is to synthesize answers from trusted clinical sources accepted in UK practice, including guidance endorsed by NICE, SIGN, and Royal Colleges, rather than to operate as a free-form conversational system detached from source material (Tytler, 25 Sep 2025).

The platform is framed as a response to two linked problems. The first is information overload: clinicians work under time pressure while navigating an expanding ecosystem of biomedical literature, national guidance, local protocols, and drug information. The second is the safety limitation of general AI tools: general-purpose LLMs may hallucinate, omit provenance, or produce recommendations that are not aligned with accepted UK practice. In this context, iatroX is positioned as a point-of-care reference layer that aims to make evidence retrieval faster, more conversational, and more usable than searching PDFs on intranets or conventional reference products such as UpToDate, DynaMed, and BMJ Best Practice.

The system is also described in public health informatics terms. By making current evidence and guidance easier to access, it may support more consistent use of evidence-based standards and reduce variation in care. This is an interpretive claim in the paper rather than a demonstrated outcome: the reported evaluation concerns adoption, usability, and perceived clinical value, not causal effects on decisions, care quality, or patient outcomes.

2. Technical architecture and safety-aware retrieval design

iatroX runs across web, iOS, and Android, with a decoupled architecture comprising a Next.js web frontend, a React Native mobile application, a Node.js/Express backend API, and MongoDB as the primary datastore. The paper further states that iatroX is registered with the UK MHRA as a Class I Medical Device, reference 2025042201417535, and that development follows structured quality and software lifecycle processes aligned with standards such as IEC 62304 (Tytler, 25 Sep 2025).

Its central technical mechanism is a proprietary retrieval-augmented generation pipeline described as a “Safety-aware RAG decision flow.” The knowledge corpus is built from publicly available UK clinical guidelines and resources from authoritative bodies, and the system runs daily monitoring scripts to detect updates and maintain currency. Documents undergo proprietary preprocessing, are split into semantically coherent chunks with average ~500 tokens and maximum ~3000 tokens, embedded using models from the GPT embedding family, and indexed in Pinecone for semantic retrieval.

When a clinician submits a natural-language question, the primary search stage retrieves up to 15 most relevant text passages. The system then computes semantic similarity between the query and retrieved passages and derives a confidence score. If confidence is low from core guidelines alone, iatroX expands retrieval to publicly available peer-reviewed research and ranks evidence by hierarchical strength, for example prioritizing meta-analyses over case studies. If confidence remains below a predefined threshold after expansion, the platform declines to answer. This refusal-to-answer behavior is explicitly presented as a hallucination-mitigation and safety feature.

If confidence is sufficient, the retrieved passages and user query are passed to a cascade of LLMs, including models from the Google Gemini and OpenAI families, together with a proprietary post-trained model called “Thea” optimized for clinical synthesis in the iatroX workflow. Answers are returned with citations and links to source material, and the figure in the paper reports a median latency of 12 seconds. The threshold logic is represented in the figure using the notation $< \tau$ and $\geq \tau$, with retrieval returning the top $k$ passages; in the text, $k$ is instantiated as up to 15 passages.

A defining property of the platform is its UK-centred content model. The retrieval corpus is built from UK clinical guidelines and resources, the answers are intended to align with accepted UK practice, and the reported user base was predominantly located in the United Kingdom. This local specificity is treated as a substantive differentiator from generic AI tools that may produce non-local recommendations.

3. Provenance, governance, and trust signaling

Provenance is central to the system’s design. The paper repeatedly emphasizes that iatroX answers are grounded in retrieved source passages, linked to citations and source links, and tied specifically to UK guidance (Tytler, 25 Sep 2025).

This provenance layer serves both epistemic and governance functions. Epistemically, it constrains answer generation to retrieved materials rather than relying on model pretraining alone. Operationally, it allows clinicians to inspect the underlying sources and verify whether a response reflects the cited guidance. In qualitative feedback, users specifically valued clear citation of UK guidelines, indicating that source linkage was not a peripheral interface feature but a primary trust mechanism.

Trust in iatroX is described as multi-component rather than reducible to answer fluency. The paper identifies several elements that users associated with reliability: MHRA Class I Medical Device registration, visible grounding in authoritative documents, UK specificity, and explicit refusal behavior when evidence support is insufficient. A plausible implication is that trust formation in clinical AI reference systems depends not only on output quality but also on visible signals of governance, localization, and bounded system behavior.

The platform is therefore differentiated from general-purpose tools such as ChatGPT or direct use of Gemini by RAG grounding, explicit provenance, UK-specific content, confidence scoring, refusal-to-answer thresholds, and regulatory framing. The paper also distinguishes iatroX from systems focused on broader medical intelligence or differential diagnosis, presenting its niche as a regulated, UK-guideline-focused clinical reference workflow with explicit retrieval safety controls.

4. Evaluation design and analytical framework

The published evaluation is a mixed-methods formative implementation/service evaluation consisting of a retrospective observational analysis of real-world usage data and a cross-sectional in-product intercept survey. The observation window covered 16 weeks, from 8 April 2025 to 31 July 2025, across web, iOS, and Android (Tytler, 25 Sep 2025).

The work was classified using the UK HRA decision tool as a service evaluation rather than research requiring REC/IRB review. The paper states that no Research Ethics Committee review was required, the lawful basis for anonymized user data processing was legitimate interests, survey participation was voluntary and anonymous, implied consent was assumed when users proceeded with the survey, and no personal identifiers or clinical data were collected. Reporting is said to adhere to STROBE and CHERRIES.

Usage analytics were obtained from Google Analytics 4 for web, Apple App Store Connect, and Google Play Console, with standard bot filtering applied. The metrics collected included unique users, engagement events, active users including DAU, WAU, and MAU, total clinical queries submitted to the RAG pipeline, and app downloads and mobile activity.

The in-product survey was designed to capture feedback in context and reduce the stronger self-selection and recall biases of retrospective external surveys. A client-side script displayed prompts to a random 10% of web sessions. Questions were single-item prompts drawn from a predefined battery and randomized across sessions and users. A browser cookie prevented showing the same prompt repeatedly to the same user. Response options were “Yes,” “No,” and “Don’t Know,” and the survey focused on usefulness, reliability, system performance, usability, and adoption intent. The study reports 1,223 total item-level responses, but because only one randomized question was shown at a time, the denominator for each specific item was much smaller.

The primary survey outcome was the proportion of positive “Yes” responses for each item. Usage analytics were summarized descriptively, and survey proportions were reported with 95% Wilson score confidence intervals. No multivariable models or inferential between-group comparisons were performed, and there was no randomized trial of the platform itself; randomization applied only to survey prompt selection and exposure. Qualitative feedback from online blogs, professional social media forums including Facebook doctor groups, and direct emails to the developer underwent rapid thematic content analysis.

5. Adoption, usage patterns, and perceived clinical value

During the 16-week evaluation window, iatroX reached 19,269 unique web users and 202,660 engagement events, corresponding to 10.5 average events per user, and the paper reports approximately ~40,000 clinical questions or queries across all platforms. Mobile uptake included 1,960 iOS downloads, while Android showed steady growth with peak daily active users exceeding 750 by late July 2025 (Tytler, 25 Sep 2025).

The platform used a usage-based registration model in which guest users could ask three free queries per week, after which registration was required. During the study period, there were 1,997 registered users and a visitor-to-registered-user conversion rate of 10.4%. The paper interprets these registrants as “product-qualified users,” suggesting that registration often followed initial evidence of utility.

Traffic sources were consistent with organic professional diffusion. The paper reports 9.1k sessions from direct traffic, 7.7k sessions from Facebook/social professional groups, and 4.7k sessions from Google organic search. These patterns are interpreted as indicating brand recall or habitual return usage, word-of-mouth growth, and spread through professional communities. The user base was predominantly in the United Kingdom.

Survey responses were generally positive, although each item had a small denominator. Reported item-level results include:

Item	Positive responses	Reported result
Useful	50/58	86.2%
Saved time today	14/23	60.9%
Would use again	14/15	93.3%
Recommend to a colleague	38/43	88.4%
Information is accurate	30/40	75.0%
Looks and feels reliable	27/34	79.4%
Speed satisfactory	45/57	78.9%
Easy to navigate	37/45	82.2%

The paper notes slight discrepancies between abstract-reported and table-reported confidence intervals for some items, but the central pattern is unchanged: perceived usefulness, reuse intent, recommendation intent, perceived accuracy, reliability, speed, and navigability were favorable. At the same time, the item-specific sample sizes were small, so the corresponding Wilson intervals were often wide.

Qualitative findings reinforce the quantitative results. Recurrent themes were speed and efficiency, usefulness in clinical workflow, reliability and governance, guideline linkage and provenance, UK specificity, and organic professional spread. Users described the tool as fast, easier than manually searching traditional guideline repositories, and valuable for rapid access to management information and synthesis of complex plans. They also explicitly linked trust to MHRA registration and visible citation of UK-specific guidance. This suggests that perceived clinical value derived from a bundle of properties rather than from conversational interaction alone: speed, localization, provenance, and governance all contributed.

6. Place in clinical workflow, limitations, and unresolved questions

iatroX is positioned as a rapid reference layer rather than a diagnostic oracle or autonomous clinical decision-maker. In workflow terms, the intended interaction is straightforward: a clinician asks a natural-language question; the system retrieves and ranks relevant UK guidance passages; if necessary it broadens retrieval to peer-reviewed research; and it either produces a grounded answer with citations and links or refuses to answer if confidence is too low, potentially prompting clarification (Tytler, 25 Sep 2025).

The paper compares this workflow with both traditional and newer alternatives. Relative to UpToDate, DynaMed, BMJ Best Practice, and guideline PDFs or intranets, iatroX is framed as faster for query-based frontline use because it allows natural-language questioning, synthesizes the answer, and links directly to the underlying guidance. Relative to direct use of general-purpose AI systems, it is differentiated by retrieval grounding, provenance, UK-specific content, confidence scoring, refusal behavior, and MHRA registration.

Several misconceptions are explicitly countered by the evaluation. First, favorable survey responses do not establish objective clinical correctness. The study measured perceived accuracy and perceived reliability, not actual answer accuracy against a gold standard. Second, adoption does not demonstrate effectiveness. Rapid uptake and positive sentiment do not establish improved workflow efficiency, decision quality, care quality, or patient outcomes. Third, the platform is not presented as a substitute for source consultation; its value proposition depends on citation linkage and source inspection rather than opaque answer generation.

The authors identify multiple limitations. The user base likely reflected self-selected early adopters. Although there were 1,223 total item-level survey responses, per-item denominators were small, including $N = 15$, 23, 34, 40, 43, 45, 57, and 58, limiting precision. The survey was anonymous and collected no professional role, specialty, grade, or other demographic data, precluding subgroup analyses. The items were single-item custom questions rather than validated instruments such as SUS or Health-ITUES. Most importantly, there was no objective audit of answer accuracy or safety against a gold standard. The paper also discloses that K.T. is the founder and lead developer of iatroX and holds equity in the operating company, which is an important potential source of bias in study design, interpretation, and reporting. Finally, the evaluation was formative and observational rather than a controlled effectiveness study.

The proposed future work follows directly from these limitations. The paper calls for objective accuracy and safety audits using standardized clinical vignettes, expert panel review, and gold-standard comparison; workflow and efficiency studies, especially time-and-motion comparisons with traditional resources such as NICE website search; validated usability evaluation with instruments such as SUS or Health-ITUES; prospective impact studies on workflow, decision quality, and consistency of evidence-based practice; and ultimately randomized controlled trials assessing effects on clinical decision-making quality and potentially patient outcomes.

Taken together, the early evidence presents iatroX as a regulated, retrieval-grounded, UK-specific clinical reference system whose distinguishing features are authoritative corpus curation, confidence-aware retrieval, refusal when support is inadequate, and answer-level provenance. The evaluation supports the narrower claim that clinicians perceived it as useful, fast, and locally relevant, and that it achieved rapid organic adoption. It does not yet support stronger claims about objective correctness, safety performance under independent audit, or downstream clinical benefit.