Halfway through the year, the conversation around artificial intelligence's role in healthcare largely has centered on clinical documentation.
Generative AI's ability to streamline charting within electronic health records has caught the attention of established EHR vendors, investors and health systems. Some AI developers and researchers are looking beyond that by preparing and piloting AI models, relying on genetic data on patients to diagnose diseases and develop treatment plans.
Related: How 4 health systems plan to adopt generative AI in 2024
“In my AI work, my favorite thing is finding people who fall through the cracks, using technology we didn’t have 10 years ago,” said Joseph Zabinski, vice president of AI and commercial strategy at precision medicine company OM1. “The point of all this is to improve diagnosis and access to care.”
OM1 uses real-world data, which is de-identified patient information derived from sources outside of clinical trials, to help clinicians predict a patient’s risk for specific diseases.
On Wednesday, Microsoft announced a partnership with Boston-based Mass Brigham General and Madison, Wisconsin-based University of Wisconsin School of Medicine to develop, test and validate AI algorithms for medical imaging analysis, which could aid in disease diagnosis. Meanwhile, Google’s DeepMind team is developing an AI research model called Articulate Medical Intelligence Explorer to help clinicians with diagnosis. The model outperformed doctors in diagnosing patients, according to a research paper Google published in January.
It is likely to take years of development and testing and potentially regulatory oversight before AI is widely used by clinicians as a diagnostic tool. A study published Wednesday from the National Institutes of Health’s National Library of Medicine showed the technology's limitations. Researchers had the AI model diagnose patients based on 207 clinical images and provide a written rationale to justify each answer. While the AI model was accurate in selecting the correct diagnosis, researchers said the model's reasoning in how it arrived at the answer was flawed.
“The AI model most often made mistakes in image comprehension, followed by reasoning and knowledge recalls — which means that the AI model made mistakes when describing the medical image and explaining its reasoning behind the diagnosis,” said Zhiyong Lu, National Library of Medicine senior investigator and author of the study, in an email. “There needs to be more research that evaluates AI in real-world scenarios.”
Here are some of areas of research into AI's use.
Dementia and Alzheimer’s Disease
The number of people living with dementia is expected to rise globally to 78 million people in 2030, from 55 million in 2020. The cost to treat Alzheimer’s disease and dementia patients is expected to balloon to nearly $1 trillion by 2050, from $345 billion last year, according to the Centers for Disease Control and Prevention.
The projections have made dementia and Alzheimer’s a prime area for research into AI's potential. Indianapolis-based Eskenazi Health has created an AI tool to identify patients likely to develop Alzheimer’s disease. Researchers at Boston University’s School of Medicine developed an algorithm that parses imaging data to determine what’s causing a person’s cognitive decline, information that can assist clinicians with diagnosis. The researchers, which published their findings in the journal Nature in July, trained the algorithm on 50,000 patients. They found clinicians who used the algorithm were 26% more accurate in their diagnosis than those who didn’t use it.
Cancer
Generative AI developer OpenAI said in June it would use a version of its GPT-4 large language model to review patient data with cancer care startup Color Health. The companies said the tool will help providers identify missing labs, imaging, or biopsy and pathology results for cancer patients.
Rochester, Minnesota-based Mayo Clinic said in October it developed an AI model that can potentially detect surgically treatable pancreatic cancer on computerized tomography imagery scans. In September, Microsoft said it was collaborating with pathology AI company Paige to create a generative AI model focused on cancer diagnosis.
The early results on AI’s ability to diagnose cancer have been promising. AI can improve the accuracy of a skin cancer diagnosis, according to an April analysis of 12 studies from researchers at Stanford Center for Digital Health. Researchers at Boston-based Massachusetts Beth Israel Deaconess Medical Center said in December that OpenAI’s GPT-4 large language model outperformed humans in diagnosing patients for breast cancer as well as other conditions.
Kidney disease, rare genetic disorders and cardiovascular care
The nonprofit American Kidney Fund said in July it was partnering with disease prediction AI company Ubie for earlier detection of kidney disease.
Researchers at David Geffen School of Medicine at UCLA and the University of California at San Francisco developed an algorithm to detect a rare genetic disorder called acute hepatic porphyria. The disease, which primarily affects women and can be life-threatening, can take up to 15 years to properly diagnose.
In cardiovascular care, researchers and clinicians at New York-Presbyterian and Columbia University developed an AI tool to detect cardiac structural abnormalities associated with heart failure on chest X-rays. In May, researchers published a study that found the AI tool outperformed radiologists who attempted the same task.
