Technology
urgent: the barrier to building has collapsed faster than the barrier to deploying safely.
Where it still goes wrong Molly Buckley, a clinical scientist in AI at Guy’s and St Thomas’, catalogued the ways these systems fail. Neural network-based models remain black boxes whose reasoning cannot be fully inspected. They learn shortcuts rather than medicine: the now-classic examples of a model distinguishing wolves from huskies by detecting snow in the background, or flagging skin cancer by spotting the ruler dermatologists place next to malignant lesions. Performance quietly degrades when a model trained on one scanner meets images from another, a problem invisible until someone looks for it. Language models hallucinate, and they do so most dangerously not at the information-retrieval stage, but at the point of treatment planning, where a fluent, confident and wrong answer is hardest to catch. And beyond the models themselves sit the human risks: the possibility of deskilling clinicians who defer to the tool, and liability questions that remain unsettled.
The clinician-builder’s journey Between capability and pitfall sat a case study of how innovation can be driven forward without a deeply technical background. Dr. Avish Arora, a maternal fetal medicine clinician and entrepreneur, described building an AI assistant for gestational diabetes management from the frontline: securing institutional review board approval, working through the regulatory determination for a non- significant-risk device study, benchmarking the system rigorously before letting it near patients, and designing a randomised controlled trial, rather than settling for testimonials. His journey demonstrated how a single,
motivated clinician can now build genuinely useful AI, while highlighting the distance
between a working prototype and defensible clinical evidence.
From prototype to patient: founders, regulators and funders on the same stage The closing session put four founders building AI for women’s health into direct conversation with the people who shape their path to deployment. Chaired by Prof. Asma Khalil, Professor of Maternal Fetal Medicine, it was a live demonstration of what each side needs from the other. The founders themselves illustrated
how varied the terrain is. Jackie Matthew, a sonographer and co-founder of FRAIYA, presented AI tools for fetal ultrasound, a modality where image quality and interpretation depend heavily on operator skill and where the workforce shortage is acute; here, AI’s promise is consistency and capacity, and the evidence bar is set by imaging regulation. Biana Schor, an AI researcher at Amsterdam
UMC, showed work applying similar techniques to gynaecology ultrasound, a field with its own diagnostic variability. Tulsi Patel, Director of Product and Technology at Hertility, addressed a different problem entirely: symptoms of gynaecological conditions such as endometriosis, where women routinely wait years for a diagnosis, and where AI-supported symptom assessment aims to compress the front end of the pathway rather than replace the clinician at the end of it. Finally, Nina Sesto PhD presented MEGI Health’s work on maternal cardiovascular monitoring, extending blood pressure surveillance into the home during and after pregnancy, where hypertensive disorders remain a leading cause of maternal death. Four companies, four intended purposes, and, as the subsequent discussion made clear, four distinct regulatory classifications, evidence
strategies and adoption pathways. The lesson for anyone evaluating such tools is that “AI for women’s health” is not a single category. An imaging adjunct, a symptom triage tool and a remote monitoring platform carry entirely different risk profiles and should be assessed accordingly.
The regulatory specialist: traceability is everything Kimberley Peven, a regulatory and clinical evaluation specialist at Scarlet, distilled the discipline that separates companies that reach market from those that stall. AI in healthcare is, in regulatory terms, software as a medical device, and everything follows from a precisely written intended purpose. The clinical claims a company makes must trace directly to that purpose, and the evidence must trace directly to the claims. Her sharpest warning was against ambition disguised as scope: a broadly scoped tool is not a stronger product but a heavier evidence burden and a larger safety liability. Narrow, well-evidenced and expandable beats broad and unproven every time from a regulatory perspective.
The regulator: from one-off assessment to life-cycle oversight Mira Jacobs, head of digital health at the MHRA and a recent arrival from the US Food and Drug Administration (FDA), offered the regulator’s direction of travel. Her central message was that AI does not sit outside regulation; it remains software as a medical device under the existing framework, with a risk-based approach to classification. What is changing is the recognition that a single pre-market assessment is no longer adequate for systems that learn and drift. She pointed to a national conversation on building a regulatory system that keeps pace with innovation while maintaining safety, and to practical sandbox programmes, including the AI Airlock and a new regional initiative pairing innovators with Trusts under controlled conditions, designed to surface real friction points and bridge what she called the valley of death between authorisation and deployment. Her framing was constructive: the structure exists, trust is earned over time, and the sandboxes exist to remove unnecessary impediments rather than to wave technology through.
Natasha Bell, Practice Development Midwife at King’s College Hospital, Ruth Bradbury, Senior Programme Manager HIN South London, and Dijana Vilic, Head of Clinical Scientific Computing at Guy’s and St Thomas’ NHS Foundation Trust
The investor: this is a growth market, judged on the same fundamentals The investor’s lens, provided by Esther Richardot Reynal de Saint-Michel of Thena Capital, was
November 2026 I
www.clinicalservicesjournal.com 49
t
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55 |
Page 56 |
Page 57 |
Page 58 |
Page 59 |
Page 60 |
Page 61 |
Page 62 |
Page 63 |
Page 64 |
Page 65 |
Page 66 |
Page 67 |
Page 68 |
Page 69 |
Page 70 |
Page 71