Lupus is a potentially debilitating autoimmune disease that can manifest in many ways, making it a challenge for doctors to diagnose and treat effectively.
To analyze this complex disease, researchers must identify distinct patterns, or “fingerprints,” within the messy, incomplete medical data found in the patient’s EHR.
An AI model developed by researchers at Vanderbilt University Medical Center uses machine learning to develop a better understanding of the underlying causes of lupus. The findings, reported in Annals of the Rheumatic Diseases, identify distinct patterns in different patients, potentially leading to more customized treatment.
“What’s particularly interesting about this study is that we can identify specific causal sources for each patient that drove the model’s recognition of the disease,” said Marco Barbero Mota, a Ph.D. student in the Department of Biomedical Informatics and first author on the study. “It is important for clinicians to understand why their particular patient was identified as having lupus or not.”
From Recognition to Comprehension
Historically, research using artificial intelligence focused on accurate recognition of the disease without consideration of its origins, said Tom Lasko, M.D., Ph.D., associate professor in the Department of Biomedical Informatics and senior researcher on the study.
“With this research we are trying to understand the underlying reasons in a causal way,” Lasko said.
“What’s particularly interesting about this study is that we can identify specific causal sources for each patient that drove the prediction.”
The research team’s machine learning model analyzed patient data without expert input, examining the interaction of different medical measurements to discover patterns on its own.
Known as Independent Component Analysis, the technique identifies independent sources that could be the cause of lupus.
Evaluating messy combinations of demographic data, laboratory results, billing codes, and medications, this approach looked for hidden patterns that acted independently of each other across thousands of patient records. This allowed researchers to narrow down 9,000 variables to 19 independent sources that signaled the likely presence of lupus in a patient record.
Unexpected Discoveries
One surprising factor that pointed toward lupus involved a rare eye condition called toxic maculopathy. While the condition itself is uncommon, researchers found that its presence in medical records along with certain medications and other disease codes helped identify lupus patients.
“A billing code for toxic maculopathy is something that an expert wouldn’t necessarily think of as one of the potential signals of lupus,” Lasko said. “It’s a second-order effect of treatment that shows how doctors are actually practicing medicine, referring patients for screening.”
This model surfaced information in data sources – like records from annual eye exams – that had not yet been explicitly connected to lupus through medical records and billing codes in the patient’s EHR data.
Implications for Patient Care
The work has valuable implications for personalized medicine. Understanding how a disease presents through analysis of different causal patterns could eventually help physicians choose more effective treatments for individual patients.
“The potential is that we should eventually be able to say: ‘This patient has lupus #128, and lupus #128 responds to this drug and not that drug,’” Lasko said. “That would help us give them the right drug the first time.”
“We should eventually be able to say this patient has lupus #128, and lupus #128 responds to this drug and not that drug.”
This approach could also improve genetic research by identifying groups of patients who share similar underlying disease mechanisms, making it easier to identify genetic factors that contribute to lupus.
The current research focused on patients of European ancestry, being based on existing data. The team is now studying how lupus presents across different groups of people. This work is particularly important, since the effects of lupus can vary with demographics.
“It’s likely that there are many other sources that might be important for other populations,” Mota noted.
The team hopes their approach will eventually help doctors understand and treat lupus more effectively across all patient groups.
