Forwarded this by a colleague? Subscribe free and get it every week.
THIS WEEK
This week: patients can buy their own breast cancer risk score, an AI can now start acne prescriptions in Utah, and a surgical overlay that surgeons say can be distracting.
In this issue:
Latest news: Breast cancer risk scores, Utah's AI prescribing pilot, a surgical AI overlay.
In the literature: An AI intake chatbot, an explainable AI tool in lung cancer, an ICU mortality model.
Spotlight: AI regulation, liability and intellectual property.
How-to: Tell me what to cover next, and forward this to one colleague.
LATEST NEWS
AI breast cancer risk scores now sell direct to patients for $249
Everlywell now offers Clairity Breast nationwide, according to the New York Times (Roni Caryn Rabin). The FDA granted De Novo authorization in 2025 for five-year risk prediction from a screening mammogram in women without known breast cancer, per its decision letter. The tool returns average (under 1.7%), intermediate (1.7% to 3%) or high (3% or more) risk. Patients pay $249 in cash, and insurance does not cover it. The company says it validated results after authorization; the Times cites no data. Critics told the Times that a low score could lead women to skip screening, and that more screening finds slow-growing cancers without proven survival benefit.
Why it matters: Expect patients to arrive with a risk score you did not order; ask what it changes about their screening, because it estimates five-year risk, not survival.
Utah lets an AI service write new acne prescriptions
Utah's Office of Artificial Intelligence Policy signed a 12-month agreement with Nolla Health, posted by the Utah Department of Commerce, letting its AI issue first-time topical acne prescriptions and renewals for adults. It covers mild-to-moderate acne and a fixed list of topical regimens, with hard stops for pregnancy, immunosuppression and severe kidney or liver disease. Physician review loosens in stages: approval of every prescription, then weekly review after the fact, then sampling of at least 10% of volume. Advancing requires no serious adverse events and at least 95% agreement, meaning the share of AI recommendations physicians accepted. The agreement cites about 96% agreement across Nolla's last 1,000 treatments, a company figure. This is one state's pilot, not a national rule.
Why it matters: If a patient says an app started their prescription, ask which state's rules applied and who reviewed it; the agreement requires public quarterly reports, so data on how it performs should follow.
A surgical AI overlay, watched in the OR
The New York Times (Vjosa Isai) visited Toronto General Hospital during a laparoscopic cholecystectomy using GoNoGoNet, an overlay from Amin Madani's group at University Health Network. Trained on laparoscopic video, it labels suggested dissection zones to help avoid bile duct injury. Madani uses it in about two cases a month inside a clinical trial that requires consent. In the case observed, the surgeons dissected without the overlay on, saying they sometimes find it distracting. The article reports no outcome data. Ethicists raised concerns about consent for data use and skill erosion.
Why it matters: Intraoperative guidance is the AI closest to the OR team; ask what a tool has shown on outcomes, not just on screen, and who decides when it is switched off.
IN THE LITERATURE
The Lancet — Brodeur et al.
Design: Prospective, single-centre, single-arm feasibility study in urgent primary care. Patients chatted with Google's AMIE from home before the visit while a physician watched live. 98 completed both steps.
Finding: No conversation met a predefined safety-stop criterion. Of 44 clinician surveys, physicians called AMIE helpful for visit preparation in 33 (75%).
Limitation: No comparison group, one site, live physician oversight, and Alphabet funded it.
Bottom line: In this 98-patient study, AMIE ran under live physician oversight without triggering a safety stop. That does not establish patient benefit or unsupervised safety.
Nature Medicine — Prelaj et al.
Design: Multicenter observational study of 2,396 patients with non-small cell lung cancer. Machine-learning models predicted immunotherapy outcomes. In a usability test, 10 lung cancer experts and 10 nonexperts predicted outcomes for 100 selected patients with and without the tool.
Finding: The clinical-and-blood model reached an AUC of up to 0.77 on the test set, beating PD-L1 and ECOG status, but fell to 0.55 to 0.72 in external validation. With the tool, sensitivity for disease control rose from 0.68 to 0.83 among experts and from 0.77 to 0.90 among nonexperts (both P about 0.02). Across all physicians, correct disease-control predictions rose 37% (odds ratio 1.37, P = 0.1). Experts followed incorrect AI suggestions more often than nonexperts (72% versus 64%).
Limitation: The external-validation drop, 20 physicians on 100 selected cases, a non-significant overall gain, and an EU-funded study with many authors reporting industry ties. Prospective validation is under way.
Bottom line: Better predictions in a small test are not better care. When a vendor quotes internal performance, remember the external-validation drop.
Anesthesiology — Mamandipoor et al.
Design: Model development and external validation on four public ICU databases: 203,434 admissions from more than 200 hospitals, 2001 to 2022. A deep-learning model used structured data from the first 24 ICU hours and, where available, notes and chest x-rays to predict in-hospital mortality.
Finding: The structured-data model had an AUROC of 0.92 (95% CI, 0.90 to 0.93) and an AUPRC of 0.53; mortality was 5.2% to 7.9%. At eight eICU hospitals the AUROC ranged from 0.84 to 0.92. In patients with notes and imaging, adding them raised the AUROC from 0.87 to 0.89 (P = 0.0167).
Limitation: Retrospective research databases, no prospective test, and the abstract reports no comparison with standard severity scores.
Bottom line: Testing across several ICU data sets is useful. There is still no prospective test or evidence that the model changes decisions or outcomes.
SPOTLIGHT
AI regulation, liability and intellectual property: what a physician panel said
At the Society for Technology in Anesthesia meeting, a panel covered AI regulation and liability, intellectual property, and device regulation. Anesthesiology News (Ethan Covey) reported it. The points apply to any physician who uses AI in care or helps build a tool. This is a trade-press account of one panel, not a legal opinion.
Regulation. Erik Smith, MD, JD, of the Keck School of Medicine at USC, described a split landscape: largely deregulatory federal policy, and states such as California and New York moving to regulate AI more directly. He also pointed to Coalition for Health AI and Joint Commission recommendations on governance, monitoring, safety-event reporting and training. Covey's read is that state and private rules may shape practice more than federal policy in the near term. Smith said direct AI use in patient care "may or may not be regulated by the FDA."
Liability. Smith argued that physicians remain responsible for AI-assisted care: an AI tool cannot practice medicine, so a physician has to be the human in the loop. The report does not say how that applies to other licensed clinicians or across jurisdictions. He urged clinicians not to wait for legal clarity and to push for safe, ethical regulation.
Intellectual property. Kenneth Holroyd, MD, of Vanderbilt's technology transfer center, called commercialization a team sport. His advice: check early whether a patent or copyright is available, ask likely buyers what they need, and choose partners carefully. Protection can shape investment and licensing options. He summed up success as people, product and money.
Device regulation. Sandy Weininger, PhD, a retired FDA engineer, said regulation both slows development and defines safe innovation. Because devices change constantly, a 1976 law created three risk tiers: novel high-risk devices need extensive evidence, while the large middle class can often reach market by showing substantial equivalence to existing products. He added that standards, which change more easily than regulations, help teach developers to think like safety engineers.
The takeaway: Before you lean on an AI tool in care this week, ask: What is its regulatory status, and does its intended use match mine? Who at your institution approved it, and who monitors it after go-live? What will your note show about how you checked its output? If you help build or pilot a tool, find advisors early, as Holroyd urged.
HOW-TO
Before you try this: check your organization's approved-tool list, and never put patient information into a tool it hasn't approved.
Thank you, and tell me what to cover next
This week's how-to is a thank-you and a request. I appreciate you reading Loading Dose AI, and I appreciate your feedback even more. Five issues in, I would rather build it from what you use than from what I guess. The survey has five questions and takes about two minutes. It asks your role, which section earns your time, whether the length works, what to cover more of or drop, and what would make you share this newsletter with friends. Every question is optional.
Open the survey. Take the five-question survey.
Be blunt. Two questions are open text. Tell me what to cover more, what to stop covering, and what would make you pass it on.
Forward this to one colleague or friend. Pick one person who would use it. Personal referrals are how a newsletter this size grows: a note from someone they trust does more than anything I can post, and every new subscriber keeps this going.
THAT'S IT FOR THIS WEEK
How was this issue?
Send us a tool or a how-to. Using something in clinic that other clinicians should know about? Submit it here and it may show up in a future issue. We ask before we publish anything and credit you by name only if you want us to.
Know one colleague who’d use this? Forward this email, or send them the link below.
Loading Dose AI is for educational purposes only and is not medical advice. Nothing here should guide the care of a specific patient, and reading it does not create a physician-patient relationship. AI tools help gather and draft each issue; I read every source and edit every word. Opinions are mine and not my employer's. Full disclaimer