T.The market for artificial intelligence in medicine is as transparent as a brick wall. New tools have achieved remarkable results in published studies. However, it is often difficult to directly compare them with similar products and determine whether they work for different types of patients.
The new company promises to change that if it can convince AI developers to subject their products to more rigorous testing.
Called Dandelion Health, the New York-based company has launched a first-of-its-kind public service that evaluates AI products based on independent data designed to eradicate weaknesses and uncover biases. raising. The company said Wednesday that its first pilot program, due to start next month, will focus on testing algorithms that use electrocardiograms to predict heart conditions.
Irving Lowe, a cardiologist and AI expert at the Ventura Heart Institute in California, said the company’s goal of becoming an AI-verified utility specifically addresses issues of racial, ethnic and geographic bias. said to be commendable. But its success, he added, depends on getting enough data from enough sources to properly assess whether a particular algorithm works across all these different groups.
“Depending on exactly which site sources you’re using and their demographics, African American and Asian populations may still be underrepresented,” said Low, who is not involved with the company. ‘ said.
Dandelion executives have anonymized data on 10 million patients from three healthcare systems: Sharp Healthcare in San Diego, Sanford Health in South Dakota, and Dallas-based Texas Health Resources. said to have collected. Sanford is the nation’s largest rural health care provider, while city-based Sharp and Texas Health Resources have concentrations of Black, Hispanic and Asian patients.
“We took great care in choosing systems that were as far apart as possible in many different dimensions of interest,” said Giad Obermeyer, co-founder and chief scientific officer of the company. He said.
Dandelion aims to fill a gaping hole in the health AI tools market by becoming an independent quality certification body. The data required for such testing is expensive, hard to come by, and there are currently no third parties to perform it. The Food and Drug Administration reviews some AI products, but its evaluation does not provide potential customers with assurances that a particular tool will work in their system or patient population. This lack of clarity makes it difficult for AI companies to win adoption and convince insurers and other stakeholders that their products are worth the cost.
Dandelion is organized as a for-profit company and has raised seed funding from three venture capital firms. The electrocardiogram algorithm is being piloted with funding from the Gordon & Betty Moore Foundation, which also supports his STAT report on artificial intelligence. The company doesn’t plan to charge AI developers during the pilot, but expects future customers, from big pharmaceutical companies to start-ups, to pay for access to its data.
The company’s co-founder and chief executive, Elliott Green, said Dandelion’s most urgent focus is gaining buy-in from a wide range of AI developers and customers interested in building trust in the usefulness of AI tools. said to get.
“This ECG validation is the beginning of the process of being able to say, ‘I don’t want these products on my patients unless they’re validated,'” he said.
Electrocardiography is an inexpensive and commonly performed test and a hot area of algorithm development. AI can analyze electrocardiogram waveforms to identify potentially fatal conditions, such as a weak heart pump or hypertrophic cardiomyopathy. Various companies have developed ECG algorithms, including Apple and medical systems such as Mayo Clinic and Cedars Sinai.
Dandelion executives said the test results would remain private unless the developers of the AI tools wanted to make them public. While this may undermine efforts to achieve transparency early on, the company’s hope is that the test will become a commonly used quality benchmark.
“It’s built into the computer science culture of public leaderboards,” Obermeier said, noting that computer vision models are often compared based on their performance on ImageNet data. “I think that’s where we’re ultimately headed.”
The company will begin accepting tests of its ECG algorithm on July 15, with an initial pilot phase expected to run for three months. It may eventually extend beyond electrocardiograms to test algorithms developed using radiographic images, clinical records, and other types of data.
This story is part of a series that explores the use of artificial intelligence in medicine and the practice of patient data exchange and analysis. This project is supported by funding from the Gordon & Betty Moore Foundation.
