This week: the MHRA puts a Centre for AI and Digital Health and adaptive rules for AI devices at the centre of its plan to 2030, three clinicians ship from an app shopfront to a pharmacovigilance dashboard, and what the FDA clearance of Epic's Deterioration Index still asks of your hospital.
Top signal
The MHRA's plan to 2030: a Centre for AI and Digital Health On 14 October the MHRA published its 2030 Strategy, and AI runs through it. Four commitments matter most to anyone building clinical software. The agency will establish a Centre for AI and Digital Health to make clearer regulatory expectations for developers of digital and AI-driven medical devices and medicines; it will create a modern, adaptive regulatory framework for AI as a medical device; it will publish an AI Medicines Development Action Plan in 2027; and it will develop an AI-enabled Clinical Trial Protocol Checker to give sponsors regulatory certainty. On the safety side it commits to using AI to analyse real-world data, and to work with the Health Data Research Service on real-time surveillance so safety signals surface sooner, alongside simpler Yellow Card reporting for patients and clinicians. It also promises a new higher-value scientific advice service that cuts advice turnaround from 90 days to 45. Read it as a strategy, not guidance: the adaptive framework and the action plan are undertakings with 2027 dates, so nothing in it is a document you can follow today. GOV.UK
Shipped this week
- Bender Apps: Edward Bender, MD, a retired cardiac surgeon, has gathered his iOS apps for surgeons, trainees and clinicians into one home, live at benderapps.dev. His post introducing it on 2 October reads in full: "Introducing https://benderapps.dev, a home for my iOS apps for surgeons, trainees, and clinicians: oral boards prep, SESATS review, cancer staging, heart team and M&M conference prep, and more." The site also lists iBronch, a bronchoscopy teaching app. It is a shopfront more than a new shelf: the App Store seller, Bender Apps L.L.C., carries EuroSCORE II from 2011 and HeartChart from 2012 alongside two 2026 arrivals, and we have seen no independent evaluation of any of them. X
- OvaTrack: Dr Petrina N. Harrison, DNP, APRN, AGCNS-BC, a board-certified adult-gerontology clinical nurse specialist at VA Palo Alto and founder of the nonprofit HopeCare Global, launched OvaTrack, a free app for women to record and communicate possible ovarian cancer symptoms, built out of her 2025 doctoral work at the University at Buffalo on diagnostic delays. "Women know when something in their bodies has changed, but too often their concerns are minimized or explained away," she said in the launch release. It is live on the App Store, released 18 September, and her own material is clear on what it is not: symptom tracking and education that does not screen for, diagnose, treat, prevent or rule out ovarian cancer, and separate from the patent-pending clinical decision support system she lists as the clinical work. App Store
- Pharmacovigilance ADR surveillance dashboard: Natalia Sofía Ortiz Prado, a pharmacist and health data specialist in Vigo, Spain, published an end-to-end pharmacovigilance dashboard over one million post-market adverse event reports, per the project's README: raw FAERS extracts cleaned in SQL, drug names standardised from brand to generic, MedDRA reaction codes mapped, then modelled in Power BI on a star schema with an executive summary page and a demographic and outcome page. Her GitHub bio reads "Building data-driven solutions for Medical Affairs." It is a public portfolio build on open FDA data rather than a deployed system, and nobody has evaluated the pipeline's accuracy, so the clinical read of it stays unproven. GitHub
Build safely
- What EHR vendors actually disclose about their AI: researchers at Georgetown University School of Medicine and the MedStar Health National Center for Human Factors in Healthcare, writing in npj Digital Medicine on 14 October, read the public AI transparency disclosures that electronic health record developers must publish for predictive decision support interventions, covering risk analysis, mitigation, governance and eight other attributes, across 35 developers. Their finding: "reporting was inconsistent and often incomplete", with substantial variation in structure and detail, and yet all 35 developers were certified as meeting the requirements. They call for standardised reporting and more rigorous certification review. It is one study of one rule at one point in time, and it is about the software your tool will sit beside rather than about tools like yours. npj Digital Medicine
- Epic's Deterioration Index is FDA-cleared, so local validation is still yours to do: Epic's Deterioration Index, version 2, won FDA clearance on 18 September (K260008) after years of use without one, and the emergency physician John Lee read the filing so the rest of us do not have to. The 510(k) summary reports an AUROC of about 0.83 for the combined target, and the filing, as Lee reads it, reports sensitivity at the medium-risk threshold of 26 per cent for Black patients, 32 per cent for Hispanic patients and 51 per cent for white patients, on a retrospective test of about 10.7 million score observations at three sites. His point is not that the model is bad but that local validation is far easier to prescribe than to do: the confounders differ at every organisation, and a willing team often cannot find the data features it would want to validate against. He suggests letting organisations adjust weights and add their own features after testing. The Epic Edge
Tools & guides
- FHIR-AgentEval: a team at Boston Children's Hospital and Harvard Medical School published, in npj Digital Medicine on 10 October, a sandbox of 43 modular tasks for benchmarking large language model agents on realistic FHIR work, appointment management and genetic testing, each task running against a resettable FHIR server with deterministic checking of both the agent's answer and the server state it produced. Across five agent configurations and four experimental settings, giving the agent long-term memory consistently improved task success and cut strategic failures such as choosing the wrong tool or confusing resource types; the best memory configuration lifted success by 9.1 per cent over baseline on held-out tasks. It is an evaluation harness for anyone building a FHIR agent, not evidence that any agent is safe to point at a real record system. npj Digital Medicine
Events
- 19 Oct (comments close): the FDA's discussion paper on how generative AI-enabled medical devices should be regulated closes for public comment on 19 October. It has been open since 18 August under docket FDA-2026-N-7874 on Regulations.gov, and the agency asks explicitly for feedback from clinicians and researchers as well as device manufacturers. fda.gov
- 16 Nov (Scotland, applications close): applications are open for the NHS Fellowship in Clinical AI, a work-based learning programme for clinically trained staff who want to build and deploy clinical AI in practice. Eight places are funded to start in August 2027: four for doctors in training, three for registered healthcare staff and accredited healthcare scientists, and one for the allied health professions. Applications close on 16 November, and prospective applicants can ask for an informal conversation first. PSD Scotland
New in the directory
- Johnson Thomas, MD (United States): an endocrinologist in Springfield, Missouri and creator of AIBx, a thyroid nodule risk tool that retrieves similar ultrasound images with known diagnoses and generates a TI-RADS report, live at thyroidbx.com and labelled there for research use only, not FDA-cleared. His public repositories include the open-source TIRADS Annotator, endocrinology calculators and MedScribe, an in-browser, on-device SOAP note generator. thyroidbx.com
- The directory now counts 88 clinician-builders with verified, evidence-graded ships, one more than last week: Johnson Thomas joined, published on 14 October, and nobody left. Browse it, forward this to a colleague who ships, and reply if you know someone who belongs in it. cliniciansthatcode.com