A newsletter on the latest in AI for healthcare.
Welcome back,
Today’s issue starts with the NHS opening the buying gate for healthcare AI. NHS SBS has launched a £900m framework to help NHS and public-sector organisations buy AI through a clearer national route.
Google and Beth Israel Deaconess Hospital researchers also brought Articulate Medical Intelligence Explorer (AMIE) into real urgent care, where patients chatted with the AI before seeing a clinician. Across 100 supervised pre-visit chats, it triggered zero safety stops and reached 90% top-7 diagnostic accuracy.
Also inside: MedGemma 1.5, Google’s medical foundation AI model, expands into CT, MRI and pathology..
Here is what you need to know today.
SUMMARY
Top Research Paper
Google and BIDMC’s prospective AMIE feasibility study tests a conversational diagnostic AI in real urgent care visits, with live physician oversight and chart-verified outcomes.
Top AI News
NHS SBS launches a £900m Healthcare AI Solutions framework to give NHS and public-sector organisations a compliant way to buy healthcare AI.
Top Model
Google’s MedGemma 1.5 expands its open medical foundation model stack into 3D CT and MRI, whole-slide pathology, localisation and clinical document understanding.
Bedside Bets
Funding, deals, and deployments across healthcare AI.
Vitea signs Hartford HealthCare to deploy its AI oversight platform system-wide: Vitea builds AI governance software for health systems; Hartford HealthCare is a Connecticut-based integrated health system that will deploy it to oversee healthcare AI use across 48,000 staff and 500+ locations. Deal value not disclosed.
Canopy partners with Illinois CancerCare to deploy oncology AI across 14 clinics: Canopy builds AI software for oncology practices; Illinois CancerCare is a community cancer care network that will use it for triage, patient communication, remote monitoring and workflow automation across 14 clinics. Canopy says its platform supports 500+ sites and can generate ~$977k annual savings per 1,000 patients.
Valar Labs raises $22m Series A to scale its AI bladder cancer test: Valar Labs builds AI tests that analyse digital pathology slides to predict cancer outcomes and treatment response; it raised $22m to scale Vesta, its bladder cancer test.
Pulse Check
Other quick reads across health AI.
Children’s Mercy partners with Basepath Health to expand cell and gene therapy delivery: Basepath Health, an AI-native operations platform for advanced therapies, will help Children’s Mercy coordinate complex cell and gene therapy workflows.
Madigan Army Medical Center’s AI-assisted pulmonary nodule registry: Madigan, a U.S. Army medical centre, is using a PHI-compliant LLM tool to prioritise pulmonary nodule follow-up.
Sentara Health scales virtual nursing and AI: Sentara, a U.S. not-for-profit health system, is using virtual nursing, ambient documentation and other AI workflows to cut documentation burden and improve patient experience.
TOP PAPER
🤖🩺 A supervised AI doctor-chat handled 100 real urgent care patients with zero safety stops
Source: arXiv · 15 March 2026
A Google and Beth Israel Deaconess team tested Articulate Medical Intelligence Explorer (AMIE), an LLM-based diagnostic chat agent built on Gemini 2.5, in a real urgent care workflow at a Boston academic primary care clinic.
One hundred adult patients chatted with AMIE up to five days before their visit. A board-certified internist monitored every conversation live through screen-share. AMIE then produced a transcript and summary for the primary care physician, while its management plans were kept for research evaluation only.
Research question
Can an LLM-based conversational AI take a pre-visit history, suggest possible diagnoses, and support urgent care safely and feasibly, with reasoning quality comparable to primary care physicians?
Source: Brodeur, Koshy et al
Approach
Prospective, single-arm, pre-registered feasibility study at Beth Israel Deaconess’s Healthcare Associates clinic from April to November 2025.
AMIE used Gemini 2.5 Pro, then switched to Gemini 2.5 Flash after 50 encounters to improve latency.
The system followed a five-phase dialogue flow: intake, history, diagnostic validation, assessment and wrap-up.
100 adult urgent care patients completed a text-chat with AMIE before their clinician visit.
Seven internist AI supervisors watched the interactions live and could stop the session using pre-defined safety criteria.
Final diagnoses were checked through eight-week chart review.
Outcomes included safety stops, conversation quality, patient attitudes, clinician feedback, and blinded comparisons of AMIE versus physician differentials and management plans.
Source: Brodeur, Koshy et al
Results
There were zero safety stops across all 100 AMIE interactions.
AMIE’s differential diagnosis included the final diagnosis in 90% of cases in the top 7 and 75% in the top 3.
Blinded reviewers found no significant difference between AMIE and physicians on overall differential quality, management appropriateness or safety.
Physicians performed better on management practicality and cost-effectiveness.
Patients became significantly more positive toward AI after using AMIE, and that improvement remained after the physician visit.
Primary care physicians rated AMIE’s summary as helpful in 75% of completed surveys and trustworthy in 64%.
Caveats
The study was small, single-centre and highly supervised.
Pregnancy, mental health and emergency complaints were excluded.
AMIE did not have EHR access, physical exam data or multimodal inputs.
Some patients could not complete the session because of device or onboarding issues.
Physicians did not always review the AMIE transcript before the visit.
Patient trust around confidentiality and honesty was lower than in earlier simulated evaluations.
Potential impact: This is a credible early template for real-world testing of patient-facing diagnostic AI. It suggests that LLM-led history taking can be introduced safely under strong human oversight, but it also shows how high the operational bar will be before these systems can move beyond feasibility studies.
TOP NEWS
NHS SBS opens a £900m route for AI companies to scale across healthcare
Source: Digital Health · 19 May 2026
NHS Shared Business Services has launched a £900m Healthcare AI Solutions framework for NHS and wider public-sector organisations.
Budget: £900m procurement framework for healthcare AI.
Scope: diagnostics, predictive analytics, robotics, operational efficiency and consultancy.
Opportunity: gives AI companies a compliant route to sell into NHS and public-sector buyers, without every trust starting from scratch.
Why it matters: AI adoption in the NHS is increasingly becoming a procurement and governance problem, not only a model performance problem. A national buying route of this size could make it easier for validated vendors to scale, while also raising the bar for evidence, compliance and contracting.
Image form NHS SBS
TOP AI MODEL
MedGemma 1.5: Google’s compact 4B model unifies 3D imaging, pathology and clinical text
Source: arXiv · Hugging Face model
Google’s MedGemma line gets a 1.5 refresh with a 4B multimodal medical foundation model. The update expands MedGemma from standard 2D medical imaging into higher-dimensional clinical data, including CT and MRI volumes, whole-slide pathology images, anatomical localisation and multi-timepoint chest X-ray analysis.
What stands out
Adds native support for CT and MRI volumes, not just 2D imaging.
Extends into whole-slide pathology, a key modality for cancer diagnostics.
Supports anatomical localisation through bounding boxes.
Improves multi-timepoint chest X-ray analysis.
Improves clinical document understanding, including lab reports and EHR-style content.
Keeps the model at 4B parameters, making it more practical for local development and research use than much larger systems.
Caveat: This is a technical report, not clinical validation. The model may be useful as a foundation for research and development, but each downstream task still needs independent testing before use near patient care.
NEWSLETTER BY:
Dr Ezekiel Dinama
MD and PhD Researcher at Cambridge University applying physics-informed ML/AI to neurophysiological research.