Special Topic: The first prescription without a doctor

Show Notes

Utah now lets software choose a patient's first prescription for acne; the terms for withdrawing the doctor's review will outlast the acne, and physicians should write them.

Read the special topic and its sources: https://physicianintheloop.org/specials/first-prescription-without-a-doctor/

The voice on this episode is generated by AI. It reads a special topic from Physician in the Loop as written; every fact in it links to its source.

What is Physician in the Loop?

How technological innovation is changing medicine for patients and the people who practice on the front lines.

This is a Physician in the Loop special topic.

The first prescription without a doctor

Utah now lets software choose a patient's first prescription for acne; the terms for withdrawing the doctor's review will outlast the acne, and physicians should write them.

On Oct. 5 a startup called Nolla Health began a pilot in Utah in which software, rather than a doctor, chooses a patient's first prescription. Adults in the state with mild to moderate acne open its app, answer a questionnaire that takes 10 to 15 minutes, scan their face and receive a treatment the software picks from eight topical formulas listed in the company's agreement with the state. The service costs $4.99 a month during the pilot, with the medicine extra. Utah's earlier agreements, with Doctronic and Legion Health, let software renew only prescriptions that a clinician wrote first.

The acne is the least important part of the arrangement. What matters is the taper, the schedule on which a doctor's review of the software's choices is withdrawn. Its terms were settled between the company and two offices of the state's commerce department, the AI office and the licensing division, and they will be read as precedent for drugs that matter far more than acne cream. Physicians should write the criteria for such tapers before regulators and vendors settle them one agreement at a time.

Utah has posted Nolla's agreement, signed Sept. 22, on its AI office's list of authorized pilots. For at least four weeks and 100 patients, two Utah-licensed physicians review each prescription before it goes out. If the software agrees with their choices at least 95% of the time, misses none of the conditions that should stop it and records no serious adverse event, it then sends prescriptions itself, and a physician reviews every case afterward, at least weekly, for at least eight weeks and until 500 patients have been treated. After that, physicians review a monthly sample of at least 10% of enrollment, and every case involving an adverse event or an escalation in full; any serious adverse event triggers a pause. The decision to advance rests with the state: the agreement says each advance will be "conducted in coordination with the State of Utah," and it bars Nolla from starting a new stage without the prior written approval of the AI office.

Oddly, the pilot that starts treatments steps back after fewer patients than the ones that only renew them. Doctronic's agreement had physicians review the renewals of its first 250 patients before anything reached a pharmacy and check those of the next 1,000 afterward, then a monthly sample of 5% to 10%; Legion Health's terms use the same numbers, counted in renewal requests rather than patients. Doctronic must also complete at least 250 renewals in each of nine drug classes before it can move on, telehealth.org reported in April. Nolla's later sample, at least 10% a month, is no smaller than Doctronic's. But it is hard to see why starting a treatment should earn independence sooner than renewing one.

What the tests measure matters as much as how long they last. Nolla says licensed clinicians agree with its treatment choices in more than 96% of real-world cases; its agreement says the figure, about 96%, comes from the last 1,000 treatments in its existing app, as judged by its current clinicians. Utah's report on Doctronic's first months gives percentages without totals: reviewing physicians agreed with its renewals 91% of the time. Agreement with a reviewer is a measure of process. And a run of 100 or even 500 patients without a serious adverse event says little about an error that strikes one patient in a thousand, which is the kind a physician's review is there to catch.

Acne is a sensible place to start. It is common and rarely dangerous, and one of its standard drugs, adapalene, was approved for sale over the counter in 2016. Nolla leaves out oral drugs and isotretinoin. Yet even creams carry an error that a review after the fact cannot undo: some topical retinoids, such as tazarotene, are contraindicated in pregnancy, and a check a week later finds that mistake after the exposure. The protection there lies in the screening questions and the drug list. Nolla's agreement leaves tazarotene off the list and stops the software when a patient reports being pregnant, trying to conceive or breastfeeding, a screen that rests on what the patient says. All eight formulas are compounded preparations that patients may buy from Nolla itself, so the software is also choosing among products its company sells.

The medical profession's formal role has been advisory. After Doctronic began, the state's Medical Licensing Board urged that it be suspended, saying it learned of the agreement only once the service was available; the state replied that the board serves in an advisory capacity and that medical professionals had reviewed the pilot before launch. The board now has a working group supporting the AI office's review of health care applications, Nolla says it worked with the board before its own launch, and Utah says it will require third-party audits of companies' claims. California, by contrast, has a new law, signed Sept. 30, that bars health facilities, clinics and physicians' offices from using AI to perform on its own any clinical function the law reserves for a licensed professional.

Medicine already knows how to hand over decisions in stages, with the terms set before the handover begins. Residents earn independence through levels that the accreditation rules define: a supervisor present or watching live, then one immediately available, then one who reviews the care afterward, each step granted by the program director and faculty. Colorado's pharmacists may prescribe birth control under a protocol approved by the state's pharmacy board, with a screening questionnaire, a fixed algorithm, the federal eligibility criteria and a process for informing the patient's own doctor. An autonomous system that screens for diabetic eye disease in primary care was authorized in 2018 after a prospective trial in 10 primary care sites, and it decides only who needs to see an eye doctor.

The strongest objection is that criteria written by physicians will protect physicians. The American Medical Association says there generally should be a human in the loop for medical decisions, and the Federation of State Medical Boards calls failing to apply human judgment to AI's output a violation of a physician's duties. Meanwhile patients with acne treat themselves or wait; Nolla says the average wait to see a dermatologist in Utah is 61 days. Caution has costs too: reviewing physicians judged nearly a third of Doctronic's escalations overly cautious. That is all fair, and it argues for criteria that can say yes. But a yes should rest on outcomes counted by someone other than the vendor, and on a taper that reverses when the numbers turn.

The principle that matters most is reversibility. Software should be handed only decisions whose worst likely mistake can be noticed and undone before it does harm; where it cannot be, as with a drug that harms a pregnancy, a human check belongs on that branch however good the average looks. The measure that matters most is what happens to patients. Harms, treatment failures, missed escalations and patients' own reports, compared with usual care, say far more than whether the software agrees with the company's reviewers, and the targets should be fixed in advance, as the eye-screening trial's were. Everything a pilot relies on, from the drug list to the numbers needed to move from one stage to the next, should be public before the first patient, as Utah's posted agreements for Doctronic and Nolla show is possible.

Just as important is who keeps the score. Someone independent of both the company and the office that approved the pilot should do the counting; Utah's auditors, vetted by its AI office, are a start. Each step back should need the approval of a named body that includes practicing physicians in the specialty. Nolla's agreement already pauses the pilot after any serious adverse event; a rise in harms, complaints or escalations short of that should put a doctor back in front of every prescription automatically, without a negotiation. And the company should carry malpractice insurance for its software's decisions, as Utah expects of the companies in its program. A specialty society, a state medical society or a hospital's medical staff could adopt such rules, and Utah's AI office, which must report to the legislature every year, is an obvious place to press them.

Acne cream is a modest beginning, and Utah has handled it with some care. But Nolla's website already lists birth control and urinary tract infections among the conditions for a women's service it marks as coming soon, and when software starts drugs like those, the terms agreed for acne will be the precedent. Physicians should write the next terms themselves, before a company and a commerce department do it for them.

This special topic was read by an AI voice. Its sources are linked at physicianintheloop.org.