Machine learning tools identify rare and undiagnosed immune diseases through patients’ electronic medical records

Researchers say machine learning tools can identify many patients with rare, undiagnosed diseases years earlier than current methods, potentially improving outcomes and reducing costs and morbidity. It states that there is. Here are the findings from UCLA Health researchers: explained in scientific translational medicine.

“Patients with rare diseases can face long delays in diagnosis and treatment, which can result in unnecessary testing, disease progression, psychological stress, and financial burden.” said Manish Butte, MD, professor of pediatrics and human genetics at UCLA. , and microbiology/immunology who care for these patients at the UCLA clinic.

“Machine learning and other artificial intelligence techniques are making inroads into healthcare. Using these tools, we have developed an approach to speed the diagnosis of undiagnosed patients. We identified patterns in medical records that were similar to patterns in patients with known infections. ”

The study focused on diseases collectively known as common variable immunodeficiency diseases (CVID). This disease often goes undiagnosed for years or decades after symptoms appear because the disease is rare, symptoms vary widely from person to person, and symptoms tend to overlap with other diseases. Common disorders.

Furthermore, although the disorder in each individual is often caused by changes in only one gene, the same genes are not changed between manifestations of the disorder, and to date more than 60 genes have been implicated. It is thought that. Without a single causal mechanism, no genetic test can provide a definitive diagnosis.

CVID is one of the most common human innate immune defects (IEIs) and is a rare disease that increases a person’s susceptibility to infection, autoimmunity, and autoinflammation. More than 500 of his IEIs have been identified, and more are discovered each year. CVID is estimated to affect 1 in 25,000 people and is associated with a deficiency in both the amount and function of antibodies and an impaired immune response.

Butte and Dr. Bogdan Pasaniuk, professor of computational medicine, human genetics, pathology and laboratory medicine at the UCLA David Geffen School of Medicine, are part of the team that developed the machine learning tool called PheNet, borrowing from the term “phenotype.” led. , “observable characteristics or characteristics of the disease found in an individual. PheNet learns phenotypic patterns from her verified CVID cases and uses this knowledge to rank patients by their likelihood of having CVID.” To do.

“The clinical manifestations of rare immunophenotypes, such as CVID, intersect with many specialties. Patients may also be seen in an otolaryngology clinic for a sinus infection. Pneumonia may also be treated in a respiratory clinic. This fragmentation of care across multiple specialists can lead to significant delays in diagnosis and treatment.” said Butte, co-senior author of the journal paper.

Even if we were able to recognize which patients had potential immune deficiencies, it would be difficult to educate all these busy professionals about immunodeficiency while hoping that they would refer those patients to us. It’s impossible. We needed to find a better way to find these patients. ”

“Our own patients report experiencing symptoms for years to decades before being referred to our immunology clinic,” Butte added. “With PheNet, dozens of patients could have been diagnosed one to four years earlier, reducing costs and improving health outcomes by getting patients into treatment years earlier. You should be able to do it.”

Because CVID does not have a single clinical symptom, it is not easy to identify the electronic medical record “signature” of this disease. The researchers developed a computational algorithm that infers EHR characteristics from disease patterns found in the records and literature of patients known to have CVID.

The software then calculates a numerical score for each patient and ranks those most likely to have CVID. Patients with high scores, who researchers describe as “hidden in the health care system,” may be candidates for referral to an immunology specialist.

Pasaniuk said the research team applied PheNet to UCLA’s electronic medical record data, which consists of millions of patient records, and followed the blind chart reviews of the top 100 patients ranked by the system. , said 74% were found to be considered likely to have CVID. . Based on these preliminary data, Butte and Pasaniuc began applying AI to the real world.

They began by validating PheNet using more than 6 million patient records from a disparate health system in the University of California data warehouse and Vanderbilt Medical Center in Tennessee. A Butte-led collaboration to bring patients identified by the algorithm to see specialists has been launched with immunology clinics at the University of California campuses in San Diego, Irvine, Davis, and San Francisco.

“We have shown that artificial intelligence algorithms like PheNet can provide clinical benefit by speeding the diagnosis of CVID, and we hope this will be applied to other rare diseases as well. ,” Pasaniuk said.

“Implementation at all five UC medical centers is already having an impact, and we are now expanding to other diseases while improving the accuracy of our approach to better identify CVID.” We also plan to teach the system to read medical notes in order to gather more information about patients and their illnesses. ”

Lead author Dr. Ruth Johnson, a former member of Pasaniuk’s lab and now a research fellow at Harvard Medical School, said the limitations of the current health care system can create tunnel vision, where different doctors see different aspects of the disease. said. But I can’t put the whole picture together. This delays diagnosis, especially for many CVID patients who present with multisystem symptoms that fluctuate over time. Artificial intelligence can overcome these obstacles.

“Every year, there is an increase in delayed diagnosis, infections, antibiotic use, emergency department visits, hospitalizations, and missed work and school days,” she says. “In addition to the financial and emotional burden this places on patients and their families, the impact on the U.S. health care system of failing to diagnose CVID in a timely manner can total millions or even billions of dollars. Probability is high.”

In addition to Butte and Pasaniuk, UCLA authors include Ruth Johnson (lead author), Alexis V. Stevens, Rachel Mester, Sergei Knyazev, Lisa A. Cohn, Marika K. Freund, Leroy Bondas, Brian L. Hill, Includes Tomer Schwartz and Noah. Zeitlen, Valerie A. Arboleda. Lisa A. Bastarache contributed from the Vanderbilt University School of Biomedical Informatics.