A newsletter on the latest in AI for healthcare.
Welcome back,
LungIMPACT is the useful kind of negative trial: nearly 100,000 primary-care chest X-rays, a real NHS pathway, and no shortcut to CT or cancer diagnosis. It asks the deployment question that is sometimes avoided: does prioritisation using AI actually improve the patient journey?
The other half of today’s issue is infrastructure. FDA Elsa 4.0, QVAC MedPsy, and MONAI point to AI becoming embedded in regulatory data, edge devices, and imaging workflows.
Here is what you need to know today.
SUMMARY
Top Research Paper
The Lung IMPACT randomised controlled trial found that AI chest X-ray worklist prioritisation shortened report turnaround but did not reduce time to CT or lung cancer diagnosis.
Top AI News
FDA’s Elsa 4.0 and HALO announcement moves internal AI closer to consolidated regulatory data and agency workflow infrastructure.
Top Model
Tether Data’s QVAC MedPsy packs text-only medical models into 1.7B and 4B variants for edge deployment, with benchmark claims that need clinical caution.
Top Repo
MONAI remains key open-source plumbing for PyTorch-based healthcare imaging, offering standardised components rather than another closed demo.
Bedside Bets
Startup rounds, deals, and moves.
Roche acquires PathAI to expand AI-driven digital pathology and companion diagnostics
Basata raised a $21 million Series A for automating healthcare administrative workflows
Pulse Check
Quick reads across health AI.
FDA Clears Rivanna’s AI Musculoskeletal Imaging System - Rivanna’s Accuro XV receives 510(k) clearance for AI-based musculoskeletal ultrasound.
UNC Health Adopts Evidently AI Platform for Clinical Data Intelligence - UNC Health scales clinical intelligence by integrating Evidently’s AI platform with existing EHR workflows.
Latest AI industry model/platform announcements - OpenAI debuts new voice models in the Realtime API while Anthropic and AWS expand enterprise features.
Focus areas for The Anthropic Institute - Anthropic’s new research arm will focus on the economic and societal impacts of frontier AI systems.
TOP PAPER
🫁AI-based Chest X-ray prioritization in the lung cancer diagnostic pathway: the LungIMPACT randomized controlled trial
Source: Nature Medicine · 24 March 2026
LungIMPACT is the largest randomised evaluation to date of AI worklist prioritisation for primary-care chest X-rays in a real NHS pathway. AI was available to reporters in both arms; only the immediate prioritisation flag was randomised. The trial answers a question that NICE has flagged as decisive for whether to recommend CXR AI products at scale.
Question
Does immediate AI prioritisation of primary-care chest X-rays reduce time to CT and time to lung cancer diagnosis, compared with the same AI being available without prioritisation?
Stock image from Unsplash (edited)
Approach
Prospective, multicentre, pragmatic RCT across five NHS Trusts (UCLH, Leicester, Nottingham, Birmingham, East Sussex and North Essex), 17 July 2023 to 31 December 2024, with follow-up to 5 June 2025.
Days were block-randomised 1:1 to AI prioritisation on or off using qXR v4.0 (Qure.ai); non-consenting design, with CXRs rather than patients as the unit of randomisation.
93,326 CXRs from 86,945 patients analysed after data cleaning (45,987 AI-on, 47,339 AI-off); 13,347 downstream CTs and 558 lung cancers.
Primary outcomes: time from CXR to CT and time from CXR to lung cancer diagnosis. Secondary outcomes spanned 2WW referrals, treatment start, stage, AI–radiology concordance and algorithm accuracy.
Results
Median time to CT was 53 days in both arms; ratio of geometric means 0.97 (95% CI 0.93–1.02, P = 0.31).
Median time to lung cancer diagnosis was 44 days (AI on) versus 46 days (AI off); ratio 0.98 (95% CI 0.83–1.16, P = 0.84).
No significant differences in time to 2WW referral (14 vs 15 days, P = 0.13), time to treatment (76 vs 72.5 days, P = 0.99) or stage at diagnosis (P = 0.34).
CXR-to-report time did fall meaningfully with prioritisation (34.1 h vs 47.0 h, P < 0.001), but did not propagate to downstream timings.
AI–radiology discordance occurred in 30.3% of CXRs; expert review found 23.9% of discordant cases had actionable findings, and 53 cancers sat in the radiology-normal/AI-abnormal group with a median diagnosis time of 106 days.
Takeaway message: The result argues against paying for worklist prioritisation as a stand-alone CXR AI feature in systems like the NHS, and reinforces that pathway redesign, not faster reports, is what actually moves diagnostic timelines.
TOP AI MODEL
QVAC MedPsy: State-of-the-Art Medical and Healthcare Language Models for Edge Devices
Source: Tether Data Research, on Hugging Face · 7 May 2026
QVAC MedPsy is a family of text-only medical LLMs at 1.7B and 4B parameters, built on Qwen3 backbones via a curriculum of supervised fine-tuning and two-stage reinforcement learning, with Baichuan-M3-235B as the sole reasoning teacher. The pitch is uncompromising edge deployment: HIPAA-aware, on-device clinical inference via the QVAC SDK and llama.cpp.
What stands out:
MedPsy-4B averages 70.54 across seven closed-ended medical benchmarks, edging MedGemma-27B-text-it (69.95) at roughly 7x smaller.
MedPsy-1.7B scores 62.62 on the same suite, beating MedGemma-1.5-4B-it by 11.42 points and matching Qwen3-4B-Thinking-2507.
On HealthBench Hard, the 4B model scores 58.00 versus 42.00 for MedGemma-27B-text-it; the 1.7B variant still scores 54.33.
Up to 3.2x token-efficiency gains versus the Qwen3 backbones, plus Q4_K_M and Q5_K_M GGUF builds that fit comfortably on phones.
Released under Apache 2.0 for research and educational use.
Caveat: The benchmarks are self-reported and rely heavily on a single LLM-as-judge family for HealthBench; real-world clinical safety, hallucination behaviour and integration into regulated workflows remain unproven, and the authors restrict use to research and education.
TOP NEWS
FDA Expands AI Capabilities and Completes Data Platform Consolidation
Source: FDA · 6 May 2026
The FDA launched Elsa 4.0, a major upgrade to its internal AI platform available to all staff from reviewers to investigators. Key additions include custom agents, automated document generation, quantitative data analysis with charts, secure web search, voice-to-text dictation, and OCR for scanned documents.
The agency also completed consolidation of more than 40 disparate data sources into the new HALO (Harmonized AI & Lifecycle Operations for Data) platform.
Elsa operates in a FedRAMP High secure environment and keeps human experts in the loop for every step.
Why it matters: By placing AI directly on top of consolidated regulatory data, the FDA is positioning itself to reduce administrative burden on scientists and accelerate the review of new therapies, setting a high bar for responsible internal AI adoption.
TOP REPO
Project-MONAI/MONAI
MONAI (Medical Open Network for AI) is the PyTorch-based open-source framework that underpins much of academic and industrial medical imaging research, sitting inside the PyTorch Ecosystem with 8.2k stars and 1.5k forks. The latest 1.5.2 release shipped in January 2026.
Useful details:
Flexible pre-processing for multi-dimensional medical imaging data, with domain-specific networks, losses and metrics.
Compositional and portable APIs designed to slot into existing PyTorch workflows.
Multi-GPU, multi-node data parallelism for training large 3D imaging models.
MONAI Bundle format and a Model Zoo for sharing reusable, packaged workflows.
Apache-2.0 licensed, with Docker images on Docker Hub and packages on PyPI and conda-forge.
Why it matters: As regulators and health systems lean harder on validated, reproducible pipelines, a maintained open-source backbone like MONAI is increasingly the default substrate for imaging AI work that has to survive clinical scrutiny.
BEDSIDE BETS
Bedside Bets
Startup rounds, deals, and moves in healthcare AI.
Roche signed a definitive merger agreement to acquire PathAI, paying USD 750M upfront plus up to USD 300M in milestones, with PathAI's AISight Image Management System set to anchor Roche Diagnostics' digital pathology and companion-diagnostics push
Basata raised a $21M Series A led by Basis Set Ventures to automate what it calls the operational layer of US healthcare
NEWSLETTER BY:
Dr Ezekiel Dinama
Medical Doctor and PhD Researcher at Cambridge University applying physics-informed ML/AI to neurophysiological research.