AI Guided a Neurosurgeon Within a Millimetre of Blindness, and It Worked

Good morning, HealthTech insiders. Two things happened recently that the profession hasn't quite figured out how to hold at the same time. Surgeons in London removed a brain tumour through a patient's nose while an AI watched the live feed and lit up the anatomy they couldn't afford to nick (blood vessels, optic nerves) in real time. The patient walked out of hospital within days, vision restored. The same week, Ezekiel Emanuel, one of the most cited medical ethicists in the country, published a JAMA paper arguing that by 2030 autonomous AI will outperform not just doctors but doctor-AI hybrids on the five tasks that define clinical medicine. His conclusion: humans in the loop will degrade AI performance.

These are not contradictory stories. The UCL surgery shows AI working as what it genuinely is right now: a trained second pair of eyes that makes a skilled surgeon safer, not a replacement. The JAMA paper is asking a harder question about what comes after that. This Sunday's signal: AI just passed its first real surgical test. And the argument about what happens next has finally moved from Vinod Khosla's keynote slides into the flagship journal of American medicine.

CLINICAL AI

In today's Signal:

• World first: UCL/UCLH AI guides neurosurgeon through live pituitary tumour removal, protecting patient's sight, real-time anatomy identification, through the nose, a millimetre of error from blindness or death (Thursday, Aug 27)

• JAMA paper: Ezekiel Emanuel and Vinod Khosla argue autonomous AI will surpass physicians and physician-AI hybrids on 5 core clinical tasks by 2030, 'humans in the loop degrade AI performance' (Friday, Aug 28)

• Rare disease AI: CHEO's ThinkRare scans EHRs to flag missed rare disease patients, 21 new diagnoses, 70% success rate, already running in clinical practice (Friday, Aug 28)

• Stroke AI: BEACON-Neuro.AI wins FDA Breakthrough Device Designation for cerebral edema prediction after acute stroke, gets into FDA's TAP pilot on first submission cycle

SURGICAL AI
1. A robot's-eye view of the human brain. UK neurosurgeons just used AI in the operating theatre for the first time.

The signal: Surgeons at the National Hospital for Neurology and Neurosurgery (NHNN), part of UCLH, removed a pituitary tumour from a 48-year-old patient while an AI analysed the live surgical video feed in real time, highlighting critical blood vessels and optic nerves. The tumour came out. The patient's sight was protected. This is a world first for live AI assistance in neurosurgery.

The details:

• Patient Rhys Hibbert, 48, from Bedfordshire had an 11mm pituitary tumour threatening his vision. The operation was performed through his nose. The AI system, developed at the UCL Hawkes Institute and running on NVIDIA's Clara IGX platform, was trained on hundreds of annotated pituitary surgery videos and identified anatomy in real time rather than relying on pre-surgical scans. Professor Hani Marcus and surgical resident Danyal Khan performed the operation.

• The trial was funded by NIHR and Google, supported by the NIHR Biomedical Research Centre at UCLH, Royal College of Surgeons, EPSRC, and Wellcome. Dr Sophia Bano (UCL Hawkes Institute) is the technical lead.

• Rhys woke with dramatically improved vision. Within a week he was walking independently without glasses or sticks.

Why it matters: A millimetre off in pituitary surgery means blindness, stroke, or death. The AI's job was not to operate; it was to see what the surgeon might miss under pressure. That distinction matters a lot right now: this is AI proving its worth as a safety net in the highest-stakes environment in medicine, before anyone has to argue about who's responsible when it gets it wrong. The regulatory and liability conversation will follow this surgery, not precede it.

CLINICAL AI
2. Ezekiel Emanuel just signed a JAMA paper arguing AI will outperform doctors on the five core tasks of medicine by 2030. The AMA disagrees. Both sides are right.

The signal: A paper in JAMA, co-authored by Penn oncologist and bioethicist Ezekiel Emanuel and venture capitalist Vinod Khosla, argues that autonomous AI will surpass both physicians alone and physician-AI hybrids on five fundamental medical tasks by 2030. Their line: 'humans in the loop degrade AI performance.

The details:

• The five tasks: taking medical histories, diagnosis, ordering tests, prescribing treatment, and managing chronic disease. The authors reviewed all AI-in-medicine studies published since January 2024. Emanuel said he spent years dismissing Khosla's thesis at conferences, but changed his mind after reading Robert Wachter's (Head of Medicine at UCSF) forthcoming book. Conflicts of interest including Vinod Khosla's AI investments are disclosed.

• AMA CEO John Whyte pushed back, pointing to a February 2026 Nature study showing most patients could not effectively interact with LLMs to extract clinical expertise in real cases without a physician intermediating.

• The de-skilling concern is the paper's sharpest edge: if trainees defer to AI for history-taking and diagnosis, they never build the baseline judgment needed to override a bad model output.

Why it matters: The paper matters less for the 2030 date and more for who signed it. When a chair of medical ethics at Penn co-authors a piece in JAMA arguing clinicians degrade AI outcomes on the core tasks of doctoring, the debate moves from VC keynotes into credentialing and liability conversations. The profession's counter-question (what exactly is a doctor being paid to do once AI handles the five core tasks) is one medicine has avoided answering properly for a decade.

DIAGNOSTICS
3. An AI scanning EHRs for missed rare disease patients is already finding people doctors hadn't thought to look for.

The signal: CHEO, the Children's Hospital of Eastern Ontario in Ottawa, has deployed an AI called ThinkRare that continuously scans EHRs for patients who may have undiagnosed rare diseases. 21 new confirmed rare disease diagnoses so far, with a 70% success rate on flagged cases.

The details:

• ThinkRare was developed by CHEO Research Institute. It scans EHR data in the background, flags cases that match rare disease patterns, and prompts genetic sequencing or specialist follow-up. A 70% positive predictive value in rare disease screening is unusually high for a first-generation tool.

• Rare disease diagnosis typically takes an average of 4-7 years from first symptom presentation, often because no single physician sees the full pattern across a patient's history. An EHR-scanning AI bypasses the 'no one thought to look' problem.

• Reported by JMIR Publications as part of a digital health feature release, August 28, 2026. are heavily favoring companies building full clinical operating layers over narrow, single-feature AI wrappers.

Why it matters: Rare disease diagnosis is one of medicine's persistent scandals, years of wrong diagnoses while the right answer sits in the chart. ThinkRare's early numbers suggest AI can find patients the current referral model structurally misses. If this holds at scale, the waiting-for-the-right-specialist model of rare disease care starts looking replaceable.

💰 FOLLOW THE MONEY

• The deals: BEACON-Neuro.AI (WashU Medicine, Dr. Raj Dhar) received FDA Breakthrough Device Designation for BEACONPredict (AI for cerebral edema prediction after stroke) on first submission cycle; raising $3.2M seed. Midi Health (CEO Joanna Strober, $1B valuation) expanding from menopause into acute and postpartum care with 550 clinicians.
  The pattern: Clinical AI capital went to tools built for the highest-stakes decision points: stroke triage, live neurosurgery, rare disease identification. The risk profile of what AI companies are willing to tackle clinically is moving up the acuity ladder. 

🩺 THE CLINICIAN’S SIGNAL

• On the UCL pituitary surgery: The surgical team described the AI as 'a second pair of eyes' rather than a decision-maker. That framing matters for liability. If the AI flags anatomy and the surgeon acts on it, who's accountable when the flag is wrong? That conversation is starting now, before the evidence base exists.

• On the Emanuel-Khosla JAMA paper: The de-skilling argument is the one worth watching in training programs. If residents defer to AI for history-taking and diagnosis during the critical learning period, they may not build the baseline judgment needed to override a bad model output. The tools may pass the five tasks; the next generation of humans may not.

• On ThinkRare at CHEO: A 70% positive predictive value on rare disease flagging is worth taking seriously. The risk is alert fatigue if the tool scales poorly. But the alternative -- the current 4-7 year average diagnostic odyssey -- is a worse baseline to defend.

🛠️ Trending HealthTech AI Tools & Startups

• ThinkRare (CHEO Research Institute): EHR-scanning rare disease flagging AI, live in clinical practice at CHEO Ottawa; 21 confirmed diagnoses, 70% hit rate on flagged cases.

• UCL Hawkes Institute Pituitary AI (NVIDIA Clara IGX): Real-time surgical anatomy recognition from live endoscopic video; trained on hundreds of annotated pituitary surgery recordings. NIHR-funded clinical trial at NHNN/UCLH.

• BEACONPredict (BEACON-Neuro.AI): AI clinical decision support for predicting malignant cerebral edema risk after acute ischemic stroke. FDA Breakthrough Device Designation granted. Not yet cleared for commercial use.

• Curai Health: AI-driven online clinic (Neal Khosla); AI handles most clinical interactions, physicians manage prescriptions and complex cases. Live commercial product.

• BEACON-Neuro.AI: Founded by Dr. Raj Dhar (WashU Medicine), AI for neurocritical care decision support, starting with cerebral edema prediction post-stroke. FDA Breakthrough Device Designation on first cycle. Raising $3.2M seed. Washington University in St. Louis spinout.

• Midi Health: Led by CEO Joanna Strober, women's telehealth unicorn ($1B valuation, $100M Series D Feb 2026). Expanding from menopause into acute and postpartum care. 550 clinicians. Triple-digit revenue growth.

• Curai Health: Led by Neal Khosla (son of Vinod Khosla), AI-first virtual clinic. Real-world experiment in autonomous AI care delivery cited in the JAMA paper.

• OmicsBank: $2.25M seed (Redesign Health, Aug 25). Clinical data infrastructure: 12.5M patient EHR records, 30M+ DICOM images, 6M pathology slides, 500K genome sequences. 90+ hospitals in Asia. Expanding to US.

📣 Highlights: Events, Lists & Deep Dives

• 📖 Trending Deep Dive: What comes after AI-assisted surgery? The regulatory and liability questions nobody is asking yet, https://thehealthtechsignal.com/p/ai-brain-surgery-what-comes-next

• 🏆 HealthTech List: The 2026 clinical AI diagnostic tools actually deployed in hospitals, ranked by evidence, https://thehealthtechsignal.com/p/top-ai-diagnostic-tools-2026

• 🔬 Diagnostic Tech: Rare disease AI (which EHR-scanning tools are live, which have evidence, and which are still demos) https://thehealthtechsignal.com/p/rare-disease-ai-ehr-tools

• 📅 Industry Calendar: September 2026 -- Sword-Headspace close (Sep 14), Maven Clinic Onward Summit (Sep 30, NYC), https://thehealthtechsignal.com/events

That's it for today!

The central question from this issue stays with me: AI just passed its first live surgical test, and a JAMA paper is now arguing the human in the loop is the bottleneck, not the safeguard. Both things can be true simultaneously, and figuring out which is which, in which situation, is the work of the next decade in HealthTech.

Before you go, was today’s Signal worth your time? Let us know. Your feedback helps us keep improving what we send you.

See you tomorrow,

Dr. Dereck Mush, MD, MBA

The HealthTech Signal is independently funded. Sponsorships help us invest in better reporting, research, and a stronger newsletter for our readers. Want to reach HealthTech founders & executives while supporting the Signal? Get in touch.

Was this email forwarded to you? Sign up here.