UCLA researchers say a new artificial intelligence tool is catching early Alzheimer’s disease that often slips past clinicians, flagging roughly four in five people the team believes would otherwise go undiagnosed. Trained on UCLA Health patient records and tuned to work more fairly across Black, Latino and Asian patients who are frequently underdiagnosed, the system is aimed at finding people earlier, when treatment options and lifestyle changes can still move the needle.
The method, called semi-supervised positive unlabeled learning (SSPUL), was detailed in a paper published Nov. 27, 2025, in npj Digital Medicine. In tests on more than 97,000 UCLA Health records, the model reached a sensitivity of roughly 77 to 81 percent across non-Hispanic white, non-Hispanic African American, Hispanic/Latino, and East Asian groups, about double the sensitivity of conventional supervised models, the authors report.
According to UCLA Health, the team baked fairness measures into training and then checked the model’s picks against genetic benchmarks. Patients flagged as likely cases but previously unlabeled showed higher polygenic risk scores and APOE ε4 counts than those the system did not flag. The release also notes that the AI looks beyond classic memory-related billing or diagnostic codes and picks up on patterns that include signals such as decubitus ulcers and palpitations, which could prompt clinicians to take a closer look and consider screening.
“We were able to capture about 80% of the people who actually would have undiagnosed Alzheimer's disease,” Dr. Timothy Chang told ABC7. He said studies estimate up to 40% of Alzheimer’s cases go undiagnosed, a gap that hits Black and Latino communities especially hard. ABC7 also highlights families like that of Ana Kelly, who recalls recognizing her mother’s dementia years before a formal diagnosis finally landed.
How the model works
The SSPUL framework learns from both labeled Alzheimer’s cases and unlabeled patient records, using race-specific probabilistic labeling and post-processing cutoffs that are tuned for group benefit equality to reduce bias, as described in npj Digital Medicine. Instead of leaning only on documented diagnostic codes, the model pulls from a wide range of electronic health record features, including diagnoses, encounter history and age, which the paper explains helps uncover likely cases that never received a formal label.
Why earlier detection matters
Earlier detection matters because disease-modifying treatments and targeted clinical referrals are now an option for people in the earliest symptomatic stages of Alzheimer’s. The Alzheimer’s Association notes that amyloid-targeting therapies such as lecanemab and donanemab are intended for early disease, which makes timely screening and specialist evaluation more consequential for patients and families. On top of that, lifestyle changes and symptom management remain key tools for slowing decline, planning care and helping families prepare for what comes next.
Limitations and next steps
Researchers stress that the tool is a flagging system, not a diagnosis, and that its output should trigger follow-up evaluation rather than replace clinical judgment. UCLA says the team plans prospective validation in partner health systems to test how well the tool generalizes and how useful it is in real-world practice before any routine rollout, and cautions that models can reveal new biases when they are used outside the environment where they were trained, according to UCLA Health. Clinicians and ethicists will also have to weigh benefits against the risk of false positives, added patient anxiety and uneven access to specialty care.
For Los Angeles patients and caregivers, the research lands as a reminder that memory concerns deserve attention, not a wait-and-see shrug. Doctors say simple screening and a careful review of medical records can be an important first step. UCLA’s team says the coming months will focus on broader testing and conversations with health systems about how to roll out the technology responsibly, if future studies support its use.
