A UCLA-led team has built an artificial intelligence tool that uses electronic health records to spot patients with undiagnosed Alzheimer’s disease, aiming to close a major care gap: widespread underdiagnosis, especially in underrepresented communities. The study, published in npj Digital Medicine, describes a model designed to promote fairness while preserving high accuracy.
Alzheimer’s and other forms of dementia have long shown disparities in diagnosis. For instance, African Americans are nearly twice as likely to have the disease as non-Hispanic whites, yet are diagnosed at a lower rate (about 1.34 times as likely). Hispanic and Latino individuals are 1.5 times more likely to have the disease but only 1.18 times as likely to receive a diagnosis.
“Alzheimer's disease is the sixth leading cause of death in the United States and affects 1 in 9 Americans aged 65 and older,” said Dr. Timothy Chang, the study’s corresponding author from UCLA Health. “The gap between who actually has the disease and who gets diagnosed is substantial, and it’s more pronounced in underrepresented communities.”
Earlier AI efforts to predict Alzheimer’s using health records relied on traditional machine learning frameworks that could miss diagnostic biases. The UCLA team’s approach uses semi-supervised positive unlabeled learning, a method crafted to balance strong performance with fairness across groups.
The researchers trained the model on electronic health records from more than 97,000 UCLA Health patients, including individuals with confirmed Alzheimer’s diagnoses and those with unknown status.
The tool demonstrated sensitivity between 77% and 81% across non-Hispanic White, non-Hispanic African American, Hispanic/Latino, and East Asian populations—substantially higher than the 39% to 53% seen with conventional supervised models.
Building on prior AI efforts that predict various diseases, the UCLA model reduces bias and disparities by examining patterns in diagnoses, age, and other clinical features. It also identified key predictors, including neurological signs like memory loss and unexpected indicators such as decubitus ulcers and heart palpitations that may hint at undiagnosed cases.
Unlike traditional models that require confirmed diagnoses for all training data, this model learns from both confirmed cases and individuals with unknown Alzheimer’s status. Fairness considerations were embedded throughout development, using population-specific criteria to minimize diagnostic gaps.
Validation involved multiple methods, including genetic data. Those flagged as likely undiagnosed cases showed higher polygenic risk scores and greater counts of the APOE ε4 allele than those predicted not to have the disease. Researchers envision clinicians using the tool to flag high-risk patients for further screening, which is especially important as new treatments become available and lifestyle interventions can slow disease progression.
Prospective validation is planned in partner health systems to test generalizability and clinical usefulness before any routine-care deployment.
“By delivering equitable predictions across populations, our model can help address substantial underdiagnosis in underrepresented groups,” Chang noted. “It has the potential to reduce disparities in Alzheimer’s diagnosis.”
Would you like this rewritten version to lean more toward a lay audience with additional real-world examples, or maintain a clinical tone suitable for medical professionals?
