UCLA’s New AI Tool Can Spot Missed Alzheimer’s Diagnoses While Reducing Racial and Ethnic Disparities

Researchers at the University of California, Los Angeles (UCLA) have developed a powerful new artificial intelligence tool designed to identify people who likely have undiagnosed Alzheimer’s disease, a long-standing and deeply concerning gap in dementia care. What makes this development especially important is that the tool was built not only to improve detection, but also to reduce racial and ethnic disparities that have persisted for decades in Alzheimer’s diagnosis.

The study was published in the peer-reviewed journal npj Digital Medicine and focuses on a problem clinicians and researchers have known about for years: a large number of people living with Alzheimer’s are never formally diagnosed, and this problem is significantly worse in underrepresented communities.

Alzheimer’s disease is currently the sixth leading cause of death in the United States and affects about one in nine Americans aged 65 and older. Despite its prevalence, many patients go undiagnosed until the disease has progressed substantially, limiting treatment options and support. According to the UCLA research team, the gap between who actually has Alzheimer’s and who receives a diagnosis is both wide and unequal.

The Diagnosis Gap and Why It Matters

Decades of research show that Alzheimer’s disease does not affect all populations equally, yet diagnosis rates fail to reflect true disease prevalence. African Americans are nearly twice as likely to develop Alzheimer’s or related dementias compared to non-Hispanic white individuals, but they are only about 1.34 times as likely to receive a formal diagnosis. Similarly, Hispanic and Latino populations are about 1.5 times more likely to have the disease but only 1.18 times more likely to be diagnosed.

These discrepancies are not minor statistical issues. Diagnosis is often the gateway to medical monitoring, access to support services, eligibility for clinical trials, and, increasingly, new Alzheimer’s treatments. Without a diagnosis, many patients and families miss out on critical care opportunities.

Traditional diagnostic approaches rely heavily on cognitive testing, specialist referrals, and patient self-reporting. These processes can be influenced by socioeconomic factors, access to healthcare, cultural differences, and implicit bias, all of which contribute to uneven outcomes.

Why Earlier AI Models Fell Short

In recent years, researchers have tried using machine learning models to predict Alzheimer’s disease using electronic health records (EHRs). While promising in theory, many of these systems were built using traditional supervised learning methods. That means they were trained only on patients with confirmed diagnoses and treated everyone else as disease-free.

This approach creates a major problem. If large numbers of people actually have Alzheimer’s but are undiagnosed, the model learns from biased labels. As a result, it may appear accurate while silently reinforcing the same disparities already present in the healthcare system.

The UCLA team recognized that to truly address underdiagnosis, they needed a model that could learn from uncertainty, not ignore it.

A Different AI Approach Designed for Fairness

To solve this, UCLA researchers developed a model based on semi-supervised positive unlabeled learning, a technique that allows the AI to learn from both confirmed Alzheimer’s cases and patients whose disease status is unknown.

Instead of assuming that “no diagnosis” means “no disease,” the model treats many patients as potentially positive but unlabeled, reflecting real-world clinical uncertainty. This is a crucial shift that aligns the technology more closely with how medicine actually works.

The researchers trained the model using electronic health records from more than 97,000 patients within the UCLA Health system. These records included demographic data, diagnoses, age, and a wide range of clinical indicators.

Crucially, the team also integrated fairness measures throughout the development process. Population-specific criteria were used to ensure that the model’s performance remained balanced across non-Hispanic white, non-Hispanic African American, Hispanic/Latino, and East Asian groups.

How Well the Model Performed

When tested, the results were striking. The new UCLA model achieved sensitivity rates between 77% and 81% across all major racial and ethnic groups studied. Sensitivity refers to the model’s ability to correctly identify patients who actually have Alzheimer’s disease.

By comparison, conventional supervised models showed sensitivity rates of only 39% to 53%, and performance varied significantly between populations. In other words, the new approach was not only more accurate overall, but also far more equitable.

The AI system identified both expected and unexpected predictors of undiagnosed Alzheimer’s. Neurological indicators such as memory loss and cognitive decline played an important role, but so did less obvious patterns. These included conditions like decubitus ulcers (pressure sores) and heart palpitations, which may signal broader health deterioration associated with advanced, unrecognized dementia.

Genetic Validation Adds Strong Evidence

To further validate the model, the researchers turned to genetic data, an unusually rigorous step for studies of this kind. Patients predicted by the AI to have undiagnosed Alzheimer’s showed significantly higher polygenic risk scores for the disease.

They also had higher counts of the APOE ε4 allele, a well-established genetic risk factor strongly associated with Alzheimer’s disease. This genetic confirmation provided compelling evidence that the AI was identifying real, biologically meaningful cases, not just statistical noise.

What This Means for Clinical Care

The potential clinical impact of this tool is substantial. Rather than replacing clinicians, the AI is designed to flag high-risk patients who may benefit from further cognitive evaluation, imaging, or specialist referral.

Early identification is becoming increasingly important as new Alzheimer’s therapies emerge and as evidence grows that lifestyle interventions can slow disease progression when applied early. Even modest delays in disease onset or progression can significantly improve quality of life for patients and caregivers.

The UCLA team emphasizes that this model is not meant to be a standalone diagnostic tool. Instead, it could function as a decision-support system, helping clinicians allocate attention and resources more effectively.

Next Steps and Future Validation

Before the tool can be used routinely in clinical practice, the researchers plan to conduct prospective validation studies in partnering health systems. These studies will test whether the model performs as well outside UCLA and whether it meaningfully improves real-world diagnosis rates.

Assessing generalizability and clinical utility is a critical step, especially when dealing with diverse healthcare environments and patient populations.

A Broader Look at Alzheimer’s and AI

This research arrives at a pivotal moment. Alzheimer’s disease remains incurable, but advances in early detection, treatment, and prevention are accelerating. At the same time, concerns about bias in medical AI systems have grown as these tools become more common.

What sets this UCLA study apart is its deliberate focus on equity alongside accuracy. Rather than treating fairness as an afterthought, the researchers embedded it directly into the model’s design.

If successful at scale, this approach could serve as a blueprint for how AI should be developed for other diseases where underdiagnosis and disparities are common, including cardiovascular disease, mental health conditions, and certain cancers.

By acknowledging uncertainty, learning from incomplete data, and actively correcting for bias, this new AI tool represents a meaningful step toward more inclusive and effective healthcare.

Research paper: https://www.nature.com/articles/s41746-025-02111-1