Researchers at UCLA Health have built a machine learning system designed to do something conventional Alzheimer’s screening tools have historically struggled with: catch undiagnosed cases in Black and Hispanic patients at close to the same rate it catches them in white patients. The work, led by Dr. Timothy Chang of UCLA’s Department of Neurology and published in the journal npj Digital Medicine, used records from more than 97,000 UCLA Health patients between the ages of 65 and 90 to train and test the model.
The Problem the Model Is Trying to Fix
Alzheimer’s diagnosis in the U.S. is unevenly distributed relative to who actually has the disease. According to figures cited in the study, African American patients are nearly twice as likely as non-Hispanic white patients to have Alzheimer’s disease but only 1.34 times as likely to receive a formal diagnosis. Hispanic and Latino patients are about 1.5 times more likely to have the disease but only 1.18 times as likely to be diagnosed. Chang described the resulting mismatch bluntly, saying the gap between who actually has the disease and who gets diagnosed “is substantial, and it’s more significant in underrepresented communities,” a pattern researchers attribute to unequal access to specialists, cultural and language barriers in cognitive testing, and clinicians’ own diagnostic blind spots.
How the Model Works Differently
The UCLA team’s approach relies on a technique called semi-supervised positive unlabeled learning, or SSPUL. Most predictive models trained on medical records treat every patient without a diagnosis code in their chart as disease-free, an assumption that quietly bakes in whatever biases already exist in who gets diagnosed. SSPUL instead allows the model to flag a subset of undiagnosed patients as likely positive cases based on clinical patterns in their records, rather than assuming an absence of a diagnosis means an absence of disease. That distinction turns out to matter a great deal for a condition like Alzheimer’s, where diagnosis itself is known to lag the underlying disease by years in many patients.
What the Model Picked Up On
Beyond the expected markers like documented memory loss, the model flagged less obvious warning signs correlated with undiagnosed Alzheimer’s, including decubitus ulcers, more commonly known as bedsores, and unexplained heart palpitations. Neither is an intuitive Alzheimer’s symptom on its own, but both can reflect the downstream effects of unrecognized cognitive decline, such as reduced mobility or missed medication management, that show up in a patient’s chart well before a formal cognitive diagnosis is entered.
The Numbers That Matter
Across the demographic groups tested, including non-Hispanic white, African American, Hispanic/Latino, and East Asian patients, the model achieved sensitivity of 77 to 81 percent, meaning it correctly identified roughly four out of five true Alzheimer’s cases within each group. Conventional models built without the bias-correction technique achieved sensitivity of only 39 to 53 percent, and importantly, performed worse specifically among the minority patients who are already underdiagnosed in real-world care, meaning conventional tools were compounding rather than correcting the disparity.
Caveats and Open Questions
The model was trained and validated entirely on UCLA Health’s own patient population, a single, large academic medical system in Los Angeles, and researchers have not yet published results showing how it performs when applied to hospital systems with different patient demographics, documentation habits, or electronic health record software. Flagging a patient as a likely undiagnosed case is also not the same as a clinical diagnosis; the tool is designed to prompt further evaluation by a physician, not to replace one. Some bioethicists have separately raised concerns about how flagged-but-unconfirmed diagnoses get handled in a patient’s record and insurance coverage before a clinician formally confirms them.
Where This Goes From Here
UCLA researchers say the next step is testing the model in prospective clinical settings, where flagged patients would actually be routed to follow-up cognitive assessments, to see whether the tool changes real-world diagnosis rates rather than just retrospective accuracy on existing records. If it holds up, the underlying SSPUL technique could plausibly extend beyond Alzheimer’s to other conditions where diagnosis is known to lag disease prevalence unevenly across racial and ethnic groups, a category that includes several chronic and neurological conditions researchers have flagged as prone to the same kind of diagnostic gap.
