Healthcare AI News 10/11/23

News

Microsoft introduces new healthcare-related tools to its Fabric analytics platform that can assemble and standardize data from multiple sources – such as EHRs, imaging systems, lab systems, medical devices, and claims data – and present it in a single view. The company also announced Azure AI Health Bot, which can answer staff questions about treatments and protocols and patient portal queries about symptoms and medical terms. Microsoft also announced a text analytics solution, along with generative AI models that create a patient history, simply medical reports into patient-friendly language, and help radiologists identify possible radiology report errors.

Google Cloud rolls out AI-powered clinician search for its Vertex AI Search platform. The company says it will speed up clinician EHR searches and to perform more complex operations such as suggesting billing codes or determining if patients meet enrollment criteria for clinical trials. It can also cite and link the source of the information that it finds to ease concerns about hallucination.

University of California health systems and other groups launch VALID AI, which will review uses, pitfalls, and best practices for using generative AI in healthcare and research. The invitation-only group’s name comes from Vision, Alignment, Learning, Implementation, and Dissemination.

Business

Permanente Medical Group will expand its pilot project with France-based autonomous scribe vendor Nabla to 10,000 physicians. The browser-based Nabla Copilot generates clinical notes from conversations between providers and patients and offers white-label API integration with EHRs.

FluidAI Medical raises $15 million in a Series A funding round. The company offers an postoperative patient monitoring tool, powered by sensors and AI, that warns surgeons of anastomotic leaks in the GI tract that has an 8% surgical incidence with 12% mortality. The company is based in Canada and its product has not been cleared for sale in the US, although it lists Cleveland Clinic and Texas Medical Center as partners.

Research

Researchers say that clinical use of AI-powered predictive models feeds data back into the EHR on which model updates are trained, which will then reduce their accuracy when their training is updated. The authors observe that retraining can actually degrade model performance, as updated EHR information disrupts the connection between presentation and outcome. They recommend that health systems document the machine learning predictions that are used on a given patient and warn that a “model-eat-model world” can render an individual model and future models worthless.

Other

Biomedical researchers tell Stat that nobody knows if a given healthcare algorithm is useful because the companies that develop them don’t share data with researchers or anyone else. The authors propose that a federal agency test algorithms against a standard test set and publish their accuracy results, including a breakout by demographic groups, that FDA would use in reviewing a product for approval. They note as a precedent NIST, which evaluates facial recognition software in publicly available reports.

Contacts

Mr. H, Lorre, Jenn, Dr. Jayne.
Get HIStalk updates.
Send news or rumors.
Contact us.