The evolution of machine learning is still in its early stages, but at the HLTH 2022 conference this week, companies shared how they are working to fine tune their approaches to AI. Those efforts included everything from improving the quality of patient data that underpin the algorithms, which has been criticized for not reflecting a diverse enough patient population, to making it easier for healthcare organizations to validate their effectiveness. Health tech companies also highlighted different approaches they are taking to how they work with providers to pilot machine learning algorithms and market them.
Dr. John Halamka, president of Mayo Clinic Platform, used his talk at HLTH to highlight an initiative to assess and reduce bias in patient data to improve the effectiveness of machine learning algorithms. Launched three years ago, Mayo Clinic Platform built an ecosystem to coordinate collaborations with health tech companies to enable innovation in healthcare.
Halamka’s talk on the “algorithmically underserved” noted that currently, when healthcare organizations use an algorithm, they often have no idea whether it performs well or not. The goal of its AI validation platform, Mayo Clinic Platform _ Validate, is to provide clinical validation for machine learning algorithms.
Mayo Clinic Platform is also partnering with other healthcare organizations to set standards and reporting for models as part of The Coalition for Health AI. In addition to Mayo, other founding members include University of California Berkeley, Duke Health, Johns Hopkins University, MITRE, Stanford Medicine, and University of California San Francisco. Industry members include Change Healthcare, Google, Microsoft, and SAS.
Halamka announced that a webinar is planned for December 7, which will offer a preview of plans for a public-private partnership to create a national register to assess the usefulness of a variety of healthcare algorithms.
“We feel we need a national set of assurance standards for algorithms,” Halamka said. The registry will host the metadata for algorithms produced in healthcare.
Health tech companies are also developing marketplaces to improve the way collaboration partners, such as providers, payers and research groups, select algorithms.
“Data may be the new oil but the data needs to be refined,” said Wavemaker Three-Sixty Health General Partner Jay Goss. One of its portfolio companies, Gradient Health, partners with medical data providers around the globe (generally hospitals and imaging centers) to curate annotated medical images for AI research labs and corporations, so that they don’t have to do one-off deals with hospitals to obtain the data. Companies can search through segmented and labeled studies, or request a custom dataset, spending less time tracking down data and more time developing new tools.
Ferrum Health developed a program to enable health systems to assess machine learning algorithms without exposing their de-identified patient data to a cloud or otherwise forcing them to centralize that data. The company, which is part of the United Healthcare Accelerator 2022 cohort, exhibited on the accelerator’s pavilion. Ferrum’s approach enables these tests to be done on-premises, behind a firewall, an approach David Miller, Ferrum’s vice president of sales – West, said is designed to de-risk their business for hospitals and health systems. The algorithms in its marketplace are FDA approved.
“We run a test of the algorithms using the hospital’s de-identified patient data to show how they perform for them,” said West. “We let our clients try it before they buy it.”
The company’s four AI Hubs include: oncology, orthopedics, cardiovascular, and breast care.
Reducing physician burnout
DeepScribe exhibited as part of the Plug and Play accelerator’s footprint at the conference. Its automated physician natural language processing software automatically summarizes a physician’s conversation with their patient and auto-populates those notes into EMR fields. Among the EMR companies it works with are athenahealth, dr chrono, AdvancedMD and Claimpower.
Earlier this year, DeepScribe closed a $30 million Series A round to support the company’s growth. The business is designed to negate the need for an in-person medical scribe, saving clinicians money.
The validation approach to algorithms seems like a natural progression in machine learning, similar to the rise of digital health apps followed by the need to validate them to ensure adoption by healthcare organizations skeptical of overhyped tech. It’s a natural progression balancing the interest in machine learning with the recognition that healthcare algorithms are not created equal.
Photo: Hemera Technologies, Getty Images