AI Helps Speed Up Rare Immune Disorder Diagnoses, Potentially Improving Outcomes

Key Takeaways:

AI-Driven Early Detection: UCLA's PheNet tool uses machine learning to detect common variable immunodeficiency (CVID) early from electronic health records.
Reduced Costs and Improved Care: Early detection with PheNet can cut unnecessary tests and treatments, lowering costs and improving patient outcomes.
Broader Implementation and Refinement: With NIH funding, the tool is expanding across medical centers and adapting to diagnose other rare diseases.

Researchers say a machine learning tool can identify many patients with rare, undiagnosed diseases years earlier, potentially improving outcomes and reducing cost and morbidity. The findings, led by researchers at UCLA Health, are described in Science Translational Medicine.

“Patients who have rare diseases may face prolonged delays in diagnosis and treatment, resulting in unnecessary testing, progressive illness, psychological stresses, and financial burdens,” said Manish Butte, MD, PhD, a UCLA professor in pediatrics, human genetics, and microbiology/immunology who cares for these patients in his clinic at UCLA. “Machine learning and other artificial intelligence methods are making their way into health care. Using these tools, we developed an approach to speed the diagnosis of undiagnosed patients by identifying patterns in their electronic health records that resemble those of patients who are known to have the disorders.”

This study focused on disorders collectively called common variable immunodeficiency (CVID), which often elude diagnosis for years or decades after symptom onset because the disorders are rare, symptoms can vary greatly from person to person, and symptoms tend to overlap with those of other, more common, disorders. Additionally, the disorders in each individual are often driven by changes in only one gene—but not the same gene from one manifestation of the disorder to another—and over 60 genes have been implicated thus far. Without a single causal mechanism, there are no genetic tests to provide a definitive diagnosis.

CVID is one of the most common human inborn errors of immunity (IEI) —rare diseases that increase a person’s susceptibility to infection, autoimmunity, and autoinflammation. More than 500 IEIs have been identified, and more are discovered each year. CVID, estimated to affect 1 in 25,000 people, is associated with antibody deficiencies—both quantity and function—and impaired immune responses.

Butte and Bogdan Pasaniuc, PhD, a professor of computational medicine, human genetics, and pathology and laboratory medicine at UCLA David Geffen School of Medicine, led a team that developed a machine learning tool called PheNet, borrowing from the term “phenotypes,” the observable characteristics or traits of a disease as seen in an individual. PheNet learns phenotypic patterns from verified CVID cases and uses this knowledge to rank patients by likelihood of having CVID.

“The clinical presentation of rare immune phenotypes such as CVID intersects with many medical specialties. Patients may be seen in ear, nose, and throat clinics for sinus infections. They may be treated in pulmonology clinics for pneumonias. This fragmentation of care across multiple specialists leads to long delays in diagnosis and treatment. It’s impossible to teach all these busy specialists about immune deficiencies with the hopes that, even if they could recognize which patients have an underlying immune defect, they’ll refer these patients to us. We had to find a better way to find these patients,” said Butte, who, with Pasaniuc, is a co-senior author of the journal article.

“Our own patients report experiencing years to decades of symptoms before they were referred to our immunology clinic,” Butte added. “With PheNet, dozens of patients could have been diagnosed one to four years earlier than they were, and by bringing patients to care years earlier, we should be able to reduce their costs and improve their health outcomes.”

Because there is no single clinical presentation for CVID, identifying an electronic health record “signature” for the disorder is not straightforward. The researchers developed their computational algorithm to infer EHR signatures from the records of patients known to have CVID and from the patterns of illnesses found in the literature. The software then computes a numerical score for each patient that rank orders the patients most likely to have CVID. Those with high scores—patients who the researchers describe as “hiding in the medical system”—would be candidates for referral to an immunology specialist.

Pasaniuc said that when the research team applied PheNet to the UCLA electronic health record data comprising millions of patient records and followed up with a blinded chart review of the top 100 patients ranked by the system, they found that 74 percent were deemed probable to have CVID. Based on these preliminary data, Butte and Pasaniuc successfully competed to receive $4 million of National Institutes of Health funding, which allows them to apply their AI in the real world.

They started by validating PheNet with more than 6 million records of patients from disparate medical systems in the University of California Data Warehouse and at Vanderbilt Medical Center in Tennessee. A collaboration led by Butte to have specialists see the patients identified by the algorithm was launched with the immunology clinics at University of California campuses in San Diego, Irvine, Davis, and San Francisco.

“We show that artificial intelligence algorithms such as PheNet can offer clinical benefits by expediting the diagnosis of CVID, and we expect this to apply to other rare diseases, as well,” Pasaniuc said. “Our implementation across all five University of California medical centers is already making an impact. We are now improving the precision of our approach to better identify CVID while expanding to other diseases. We will also plan to teach the system to read medical notes to glean even more information about patients and their illnesses.”

Lead author Ruth Johnson, PhD, a former member of the Pasaniuc Lab who now is a fellow at Harvard Medical School, said limitations of the current healthcare system can result in tunnel vision, where different doctors see different aspects of a disease but are unable to put the whole picture together. This delays diagnosis, especially for the many CVID patients who have multisystem manifestations that fluctuate over time. Artificial intelligence can overcome these obstacles.

“For every year a diagnosis is delayed, there is an increase in infections, antibiotic use, emergency room visits, hospitalizations, and missed days of work and school,” she said. “In addition to the financial and emotional toll this takes on patients and their families, the aggregate impact to the US health system of failing to diagnose CVID in a timely fashion is likely to be in the millions or billions of dollars.”

- This press release was originally published on the UCLA Health website and has been edited for style and clarity

AI Use Disclaimer: Parts of this text, such as the title and key takeaways, may have been generated using AI. All AI-generated content is fact-checked and edited for clarity.