|
3, 2, 1: Health AI Brief
Every Friday
July 17, 2026
|
|
AI is reshaping healthcare fast. Below are 3 key AI developments, 2 studies, and 1 takeaway for this week to help you better lead with AI. Target read time: 5 minutes. |
|
3
Market Signals
Preventive‑health startup, Neko Health, closed a $700 million Series C led by Lightspeed valuing the company near $7 billion, up from $1.8 billion just 18 months ago. The 60‑minute scan pairs proprietary sensors for skin, cardiovascular, and circulatory irregularities with on‑site bloodwork and Apple Health data. More than 100,000 people have been scanned across the UK and Sweden, 350,000 sit on the waitlist, and 75% of members book next year's scan before leaving their first. The round funds a Manhattan flagship later this year. My Take
$700M is impressive but it's backed by 100,000 people scanned and a 75% rebook rate. Surprising given people are paying out of pocket. Key question will be who pays the cost of chasing down incidental findings? CMS's proposed 2027 Physician Fee Schedule, out this week, would end Medicare payment for remote monitoring delivered by third‑party vendors: only clinical staff employed directly by the billing practice would qualify. It also adds a separate initiating visit before monitoring starts and limits remote therapeutic monitoring to established patients. Trigger is an OIG review that found ~43% of enrollees weren't receiving all 3 required service components, alongside billing for services never performed. Comments close September 14; changes would take effect January 1, 2027. My Take
Last week Medicare tied chronic‑care tech payment to outcomes; this week it pushed remote patient monitoring back inside the doctor's practice. Through all of this, CMS is ultimately clarifying who is responsible, and paying accordingly. UC San Diego surgeons teleoperated general‑purpose humanoid robots, 5 feet tall and 60 pounds each, through laparoscopic gallbladder removals in live pigs, published in Nature. One procedure ran with 2 robots on the sterile field and no human surgeon beside the table. The team is candid about the gap: the robots needed recalibration mid‑surgery, and the procedures took much longer than with purpose‑built surgical systems. My Take
Robotic surgery has been around for a really long time. This is pushing towards that frontier where robots operate solo. Still early, and very hard to imagine, but not something to rule out. |
|
2
Research Studies
Frontier AI models underperform on neuroimaging for a structural reason: the scans they would need to learn from sit inside health systems. NeuroVFM was trained on 5.24 million uncurated MRI and CT volumes generated during routine care, a paradigm the authors call "health system learning." Paired with open‑source language models it beat frontier models on accuracy, triage, and expert preference, with fewer hallucinated findings. My Take
Every health system sits on years of imaging it treats as a storage cost. This paper reframes that archive as what it is: training data the frontier labs can't get (without the right partnerships and permissions). Researchers benchmarked 4 consumer AI models (GPT‑4o‑mini, Gemini 2.0 Flash, Claude 3.5 Haiku, and Llama 3.1) against 600 pediatric health queries: 300 real caregiver questions and 300 adversarial variants built on 6 pressure tactics. The models stayed safety‑appropriate 95.5% of the time, and pressure made them more careful, not less; safety scores rose under adversarial conditions across all 4 models. The one pattern that reliably eroded boundaries: claiming false medical expertise. My Take
Parents already ask chatbots about fevers at 2 am, so 95.5% is genuinely reassuring but still not ideal at scale. Things will continue to improve. Interestingly, just a better, safety‑focused prompt lifted safe responses by nearly 6 pct points. |
|
1
Key Insight
The moat is no longer the code
US digital health raised $7.4 billion across 244 deals in the first half of 2026, up $1 billion from the same period last year. The concentration is the real story: 20 megadeals took 45% of every dollar invested, roughly double their 2024 share, while 115 digital health companies were acquired, 71 in Q2 alone, the busiest M&A quarter since 2021. Rock Health's explanation for who wins: "As AI lowers the barriers to building, founders with deep experience inside the healthcare organizations they're selling to often have a clearer view of where the most meaningful (and solvable) problems exist." The report names 4 moats: founder domain edge, owning operating layers, hands‑on delivery, and network effects. None of them is the model. Takeaway
In a world where you can build anything (faster, cheaper, and even better), investors are making clear the moat isn't the building. It's deeply knowing healthcare (from the inside) to sell efficiently and deliver measurable results with a positive ROI. |
|
Know someone who'd find this useful? Share |
Humanoid robots just performed live surgery
3 signals, 2 studies, 1 insight for health leaders
Jul 18, 2026
•
3 min read