AI therapy apps outrun US regulators as health survey drops disability questions

The proposed redesign of the National Health Interview Survey, disclosed on 1 October, would cut the adult questionnaire from 482 questions to 150, according to STAT. Items on hearing aids, fatigue, cognition and mobility supports such as wheelchairs and scooters are gone.

That matters for anyone trying to judge whether an AI therapy app works. The survey, run by the Centers for Disease Control and Prevention's National Center for Health Statistics, has collected population health data since 1957. A recent paper co-authored by CDC statisticians estimated that past surveys may have missed 75% of people with intellectual and developmental disabilities, STAT reported. Removing questions makes the denominator worse, not better.

"You're making us choose between bad and worse," Scott Landes, a sociology professor at Syracuse University, told STAT.

Katy Neas, chief executive of the advocacy group The Arc, called the change a "head-scratcher." "If we don't have the government as a trusted source of data, asking questions about real people and what their real experiences are, we're never going to get better [health outcomes]," she said. A Department of Health and Human Services spokesperson declined to explain why the disability questions were struck but pointed to a push to lower data collection costs, STAT reported.

A rulebook that arrives late

Mental health apps have no equivalent of the National Health Interview Survey: no single instrument that tracks who uses them, who is harmed and who benefits. State legislatures, professional bodies and, increasingly, the courts have filled the vacuum. Pennsylvania lawmakers voted on 1 October to regulate AI use in healthcare, according to Patch. The bill's specifics are not detailed in the reporting, but the vote signals that states are no longer waiting for federal guidance.

That guidance, when it exists, moves slowly. The Federal Reserve Board asked for public comment on two stablecoin proposals on 24 September, with a 60 day window after Federal Register publication. The subject is payment stablecoins under the GENIUS Act, not mental health, but the shape is familiar: a new technology arrives, agencies consult, and rules follow years later.

In the meantime, clinicians and researchers are publishing their own evidence base. A small pilot study from UC San Francisco, published 23 September in JAMA Psychiatry and funded by the National Institutes of Health, compared depression symptoms in the same 20 participants before and after they moved from their usual ultra-processed diet to healthier food. Participants cut ultra-processed intake by more than 85% and showed moderate to large improvements across three measures of depression, with anxiety scores improving modestly, according to ScienceDaily's account of the study.

Senior author Andrew Krystal said the pilot "successfully established the viability of carrying out this research in a larger-scale trial." Corresponding author Nyasha Chagwedera pointed to a "leaky, out-of-balance gut" as one possible mechanism: "This gut damage can create low-grade, chronic inflammation throughout the body that either directly causes depression or makes people more vulnerable to it."

The study is preliminary. Twenty people, average age 44, 70% female, all with a high BMI and at least one metabolic disorder. It is not evidence that an app can substitute for a dietitian, and no one involved claims otherwise.

Where the money is going

Money, however, is moving toward automation. Parakeet Health raised $10 million to use AI to call, text and email patients, Endpoints News reported exclusively on 1 October. Devoted Health, a Medicare Advantage insurer and medical group, announced nearly $1.2 billion in fresh funding the same day after enrolling hundreds of thousands of new senior members, according to Endpoints.

The same week, the Blue Cross Blue Shield Association estimated that hospitals' use of AI-assisted medical coding contributed close to $1 billion, or $942 million, in additional costs for its health plans between 2023 and 2025. BCBSA said much of the increase came from secondary diagnoses that moved patients into higher-paying reimbursement categories. Luke Chalker, BCBSA's senior vice president of product and data science, told CNBC that roughly 70%, or $653 million, of the billing it identified was tied to additional diagnoses not accompanied by a change in care.

Chalker stopped short of attributing the entire increase to AI. "While multiple factors contribute to coding intensity, the findings suggest AI-enabled coding and documentation tools are playing a role," he said. Christopher Whaley, a health economist at Brown University, told CNBC that AI appears to be "accelerating, and in some sense making it easier to capture, the existing and underlying billing incentives that are in the system." He added that in many cases the additional diagnoses are legitimate and were simply not captured before.

Endpoints framed the same evidence bluntly on 1 October: "For all the talk about AI's potential to cut healthcare costs, it seems to be doing the opposite."

None of this is mental health app regulation in the narrow sense. It is the financial and administrative substrate on which any such regulation will sit. If coding tools can move a patient into a higher-paying category, an app that generates billable interactions has the same incentive.

Science turns to sex differences

The journal Science devoted a special issue to women's health, published on 1 October, edited by an all-female team: Yevgeniya Nusinovich, Sacha Vignieri, Joana Osório and Sarah Ross. The editors named political attacks on women's rights and healthcare, plus funding cuts to research in the area, as urgent motives. "The importance of doing clinical trials specifically in women has gained recognition over the years, and it's something that should continue to be elevated," they said, according to STAT.

One review in the issue covers sex-specific mechanisms of pain, including the role of sex hormones, immune-neural interactions, the gut microbiome and ovarian hormones. "Women are desperately waiting for concrete answers that might help eliminate the common 'all in your head' approach to their illnesses," the authors wrote. Another argues that XX and XY chromosomes should always be considered in precision medicine strategies.

That last point cuts directly at digital mental health. An app trained on skewed data, deployed to a population whose pain has historically been dismissed, is a precision medicine problem before it is a software problem. Nusinovich, a senior editor at Science, told STAT that post-reproductive health research differs from reproductive research because "the focus here is not on improving reproductive success but on considerations for a woman's own health."

Regulators have been slow to catch up with the apps themselves. That gap has been documented repeatedly, including in pieces on how US regulators struggle with AI therapy apps and why regulators cannot keep pace with AI mental health apps filling provider gaps. The general finding, across those accounts, is that the tools proliferate faster than the oversight.

China offers a contrasting model. Rest of World reported on 24 September that Chinese patients book hospital appointments through WeChat or Alipay and can chat with an AI avatar of a doctor about insomnia or diabetes. Ruby Wang, a physician and consultant, told the outlet that in the UK and US, "tech is always an add-on in the health system. You are trying to screw an extra bit onto a rusty machine, whereas in China, because the infrastructure and the digital rails are so technologically enabled, society accepts that as normal." She added that China's digital health is not directly exportable: "the West would never accept an AI doctor avatar."

What comes next

The Federal Reserve's stablecoin comment period closes 60 days after Federal Register publication, a reminder that even well-defined financial technology takes months to rule on. Mental health apps have no comparable timetable.

Researchers at University College London and Stanford, among others, published a framework on 28 September for learning continuous patient trajectories from electronic health records. The paper, arXiv:2609.36144, describes EHRFlow, a model trained on real-world clinical datasets covering more than one million patients, including an independent external validation cohort. In a controlled counterfactual simulation, conditional guidance approximated the known effect of an antihypertensive intervention without training a task-specific outcome model.

That is a technical result about forecasting, not a product. But it points to where the regulatory question is heading: models that infer what would have happened to a patient under a different treatment are precisely the tools a regulator would need to audit an AI therapy app. The methods are arriving. The survey questions are leaving. And the committees that would write the rules are about to change hands.

At least five Republicans on the Senate finance committee, out of 14, are set to leave, STAT reported on 1 October. Both the finance and health committees face major turnover after the midterms. Whoever inherits them will decide how much of the AI health stack gets regulated, and how much of the evidence base survives to inform it.