The latest in Epic data lawsuit

Brittany Trang writes: In a legal filing last week, GuardDog Telehealth admitted to requesting patient health records under the guise of "treatment," while feeding the records to law firms. Epic sued GuardDog over the practice earlier this year in its suit against Health Gorilla. A judge must still accept the stipulated judgement including an agreement that GuardDog will never again use TEFCA or Carequality, the information exchange where GuardDog obtained health records.

While it's hard to tell how this will influence the eventual outcome of the lawsuit, Epic considers it proof that its concerns are real and that policymakers should take it seriously.

"I think now it's time for ASTP to say how they feel," an Epic spokesperson told me. "Now it is time for ASTP to say what steps they are taking to ensure the integrity of the federal directory."

In other Epic lawsuit news, at least five suits were filed against the company last week. Three of them were similar class action suits that, in an interesting twist, allege Epic improperly disclosed patient records in the "data breach" that Epic is suing others over. A class action lawsuit was also filed against Health Gorilla for its negligence in not vetting the parties it onboarded to its network. If you have a view on how all of this will shake out, reach me: brittany.trang@statnews.com

artificial intelligence

Nvidia's new radiology partnership

Katie Palmer writes: Radiology AI is quickly moving away from point solutions. No longer is it enough for an algorithm to simply pick out a pneumothorax from a chest X-ray, or generate text findings from an image. Increasingly, an imaging model needs to contribute to a clinician's reasoning.

With those needs in mind, Hoppr has added two Nvidia open models to its imaging model platform, which allows customers — from AI device companies to individual radiology practices — to build and deploy new AI tools. One creates synthetic images, while the other introduces reasoning capabilities. "This fills the gaps that we had in our models," said Hoppr CEO Khan Siddiqui. "It gives our customers the ability to build much more interesting solutions."

Like what? Synthetic images could be used to build a training dataset for a rare pathology result, said Siddiqui. And one user told him they want to use the appearance of lung nodules to predict cancer staging, which could be supported by reasoning models when they incorporate genetic and outcomes data for patients.

Expect to see plenty more health AI news out of Nvidia's GTC conference this week.

Research

Brain implant helps paralyzed people QWERTY style

A new study showed two people with paralysis could reliably use a brain-computer interface to simulate two-handed typing on a QWERTY keyboard. Brain implants have used various methods to help people communicate, including typing letters with eye-tracking or simply by thinking about them.

As Rose Broderick reports, the new approach takes advantage of users' often deeply ingrained familiarity with the standard keyboard. In the study from researchers at Brown University and Mass General Brigham, a person with ALS was able to type up to 110 characters per minute with 95% accuracy.

large language models

Bot talk: Microsoft, Verily, OpenAI, Limbic

Microsoft joined the army of companies practically begging you to upload your health records to their AI services. Last week, the company announced Copilot Health, its latest swing at a consumer bot that answers users' health questions. As some have noted, Microsoft already launched another version, called Copilot for Health, in the fall. The company says you can disconnect your health records from the service at any time and that "your information in Copilot Health is not used for model training." The new offering will compete with OpenAI's ChatGPT Health, which has name recognition on its side — for now.
Speaking of ChatGPT Health, OpenAI's Karan Singhal on Friday responded to a recent Nature Medicine paper that found the company's chatbot did not respond appropriately to medical emergencies. The paper has been held up as an example of the risk posed by OpenAI's aggressive move into offering consumer health information. Singhal argues that the study did not allow the bot to ask followup questions and thus did not reflect real-world performance of the bot. OpenAI re-ran the test using the researcher's prompts and found that in "over 80% of the apparently concerning cases, the model asked for additional context so it could produce a better informed response."
Elsewhere, Verily announced general availability of Verily Me, its free service that allows anyone to get a clinician review of their health records and also includes a chatbot, called Violet, that can answer questions about your records. A new feature allows Violet to have a multi-turn conversation with users about symptoms they are experiencing. One thing to note before you hand over your medical history: Verily Me's HIPAA authorization allows the company to use your personal health information to "develop, improve, or offer services, products, and technologies related to health and wellness" and to disclose your PHI to "health and wellness companies that provide products, services, and technologies that may be of interest to you."
Mental health AI company Limbic just published a multi-faceted Nature Medicine paper that has gotten quite a bit of attention. In the most interesting part, 234 people were randomized to engage in "therapy-like" interactions with text agents powered behind the scenes by either human therapists, a standalone foundation model (e.g. GPT-4), or foundation models using Limbic's "cognitive layer" on top. The conversations were evaluated for therapeutic quality by human raters. Agents using Limbic's cognitive layer outperformed human clinicians and the standalone LLMs. It's worth noting that these weren't real clinical conversations. Participants were told the purpose of the study was to evaluate therapy agents and were told to make up a scenario to discuss.

Health tech news roundup

Maven Clinic, an employer-focused women's health and fertility company, announced it would offer its services to anyone who can pay.
Turquoise Health, a pricing data, contracting, and revenue cycle platform, announced it raised a $40 million Series C Round led by Oak HC/FT with participation from Andreessen Horowitz, Adams Street Partners, and Yosemite.
Conduit Health, developer of a tech platform that helps patients to get durable medical equipment at home, raised a $17 million Series A led by Drive Capital.
Nadia Care, a maternal health company that works with Medicaid plans, raised $12 million led by Valtruis.
Qualified Health, a company helping health systems with AI, is raising about $100 million from NEA and SignalFire, Axios reports.

More around STAT

Check out more exclusive coverage with a STAT+ subscription

Read premium in-depth biotech, pharma, policy, and life science coverage and analysis with all of our STAT+ articles.

What we're reading

For decades, they've set the record straight in biology. Next up: science's reproducibility crisis, STAT
New technologies promise to transform organ transplantation. But when?, STAT
What publishing leaders say about AI when they're not on panels: A pulse on all things AI, Scholarly Kitchen
Intuitive Surgical hit by phishing incident, Medtech Dive