AI Model Shows Promise for Improving Dementia Diagnosis
July 8, 2025
Using data from the National Alzheimer’s Coordinating Center (NACC), along with several other Alzheimer’s disease databases, a multinational team of researchers is building an artificial intelligence (AI) model that could significantly improve diagnosis and monitoring of Alzheimer’s disease and other forms of dementia. The work, described in the journal Nature Medicine last July, used multiple strategies to validate the computer model, with an eye toward using AI to augment physicians’ own diagnostic skills.
Vijaya Kolachalama, PhD, Associate Professor of Computational Biomedicine at Boston University’s Alzheimer’s Disease Research Center in Boston, MA, led the project, working with a large team of collaborators from around the world. “It’s a multidisciplinary team, and we are trying to put these people together … to see if we can address some really important questions,” says Kolachalama. Besides the computer scientists, graduate students, and medical students in his own lab and others, he enlisted help from practicing neurologists who are experts in diagnosing dementia.
“For diagnosing these conditions, you need experts, these behavioral neurologists, and there are very, very few in this world who have the right skills and experience to properly diagnose them,” says Kolachalama. Besides the need for highly skilled and specialized physicians, dementia diagnosis presents layers of additional challenges. Dementia itself can stem from multiple causes, and in many patients they overlap, ranging from Alzheimer’s disease and Parkinson’s disease to depression, anxiety, and vitamin deficiency. At the same time, early diagnosis is crucial not only for developing new therapies, but also for helping current patients. “Unfortunately, for most patients, if they have been diagnosed with moderate to severe dementia, it might be too late, so you want to do these things early on,” says Kolachalama. The development of this AI diagnostic tool is particularly timely given recent FDA approvals of new Alzheimer's treatments, which require precise diagnosis of specific dementia types for appropriate patient selection.
While many researchers are focusing on using blood tests and brain scans to try to identify the early stages of dementia, those approaches alone won’t provide a complete solution. “I think that’s amazing, and we need those things, but in reality a single blood test is not going to completely solve the problem,” says Kolachalama. He gives the example of cholesterol testing, a standard part of primary care practice. While high cholesterol can certainly indicate a range of possible problems, that measure alone doesn’t reveal what’s causing it, or point to the best treatments. “So, blood tests are definitely useful for getting a good idea on one dimension of a problem, but I think diagnosing the root causes of dementia is a lot more complicated than just a single blood test or a single [imaging scan],” says Kolachalama.
For most patients, the process of arriving at a dementia diagnosis involves multiple steps and spans at least several months. Family members might observe an older relative behaving strangely or forgetting things and eventually decide to take them to a primary care doctor. That, in turn, may lead to a referral to a neurologist for a full battery of tests. “If you really think about what the neurologist is doing, they are taking all these different forms of data, going back in time to what happened [at the primary care visit], collecting all this information and then trying to make a decision,” says Kolachalama. He adds that “effectively the neurologist is a data scientist.” With that in mind, he and his colleagues set out to create a tool that could help neurologists manage the available data and reach more accurate diagnoses.
Kolachalama emphasizes that the NACC database in particular was and remains crucial for this work. “Without that database, I don’t think we would’ve been able to be here,” he says. He adds that besides providing a massive quantity of data, the NACC has also set the standard for structuring it: “The NACC has actually served as a template for us, which I think is extraordinary.”
Even with a well-curated database, though, the first challenge in building a diagnostic model is that the incoming data aren’t always complete. For example, a patient with a pacemaker wouldn’t be able to get a magnetic resonance imaging (MRI) scan, while another patient might refuse to have their blood drawn. “We are agnostic as to whether the patient had an MRI or a blood test or some specific neuropsychological exam, we are just saying okay, give us whatever is available … and then we are going to try to make sure that the AI actually learns from these different combinations and variations in data,” says Kolachalama.
Typically, AI researchers handle missing data by imputation, a statistical technique that replaces the missing data with substituted values inferred from other data. Kolachalama’s team rejected that approach, instead insisting that their AI model learn from incomplete patient records without trying to fill in the blanks. This shift in approach was one of the project’s major innovations.
In order to build such a flexible AI, though, the researchers needed to feed it data from enormous numbers of patients with different medical histories, ethnic backgrounds, and demographics. “The NACC was the big one, because it’s basically an amalgamation of data collected across 36 different Alzheimer’s disease research centers in the US, so it’s a pretty good, diverse data set, and very comprehensive,” says Kolachalama. In addition, the team mined data from eight other databases from patients in several countries, ultimately drawing on more than 50,000 participants.
Besides examining diverse populations and different types of clinical information, the investigators also needed data they could separate into training and testing sets. “When we train these models on, say, cohorts number one through five, we want to make sure we can evaluate the performance of these models on the remaining cohorts that the model has not seen before,” says Kolachalama.
Indeed, a majority of the work published in the paper involves multiple levels of validation, a process Kolachalama says took more than a year. Looking at the AI’s ability to diagnose dementia in patient cohorts it hadn’t seen before was only the first level of validation. In addition, the scientists compared the diagnoses predicted by the model to biomarkers for dementia from a subset of patients. For example, some patients had participated in clinical trials in which they got positron emission tomography (PET) scans. “It’s not done on every patient, because it costs $8,000, but some research participants had that biomarker data, which means we can actually go and see in their brains where the amyloid deposits or any other biomarkers actually were present,” says Kolachalama. The AI model didn’t use PET scans as an input, but its predictions nonetheless aligned with the scan results in patients who had had them.
In about 150 cases, the team also had access to pathologists’ reports on postmortem brain samples, long considered the definitive way to diagnose Alzheimer’s disease and other dementias. Once again, the AI predictions aligned tightly with the clinical reports. Finally, the investigators enlisted a team of neurologists, who evaluated about 100 of the patients from the databases. Those analyses compared the accuracy of neurologists alone, to a neurologist’s diagnosis paired with an AI prediction. “When the AI’s predictions plus the neurologist’s predictions were combined, the accuracy of that was about 26% higher on average than just the neurologist alone,” says Kolachalama. He emphasizes that the goal has never been to replace neurologists, but to augment their capabilities.
One of the most significant breakthroughs in this research addresses a reality that has long complicated dementia diagnosis: many patients don’t have just one cause for their cognitive decline. In real-world clinical practice, it's common for patients to have multiple overlapping conditions that contribute to their symptoms. “This AI can actually predict if somebody has two causes, two reasons why they have dementia, like Alzheimer’s and vascular pathology, or Parkinson’s and depression, and so on … which has never been done before,” says Kolachalama. Previous AI diagnostic tools were designed to identify single conditions, forcing doctors to choose one primary diagnosis even when multiple factors were clearly at play.
With the initial research and validation results published, Kolachalama has been working to move the technology into clinical trials. “Clearly we want to build a tool that can be assistive, not just in neurology practices, but in some other settings as well,” he says. For example, one collaborating site is the Carle Foundation Hospital in Urbana, IL, part of the University of Illinois system. There, geriatricians see patients for annual wellness visits, similar to basic health checkups. The hospital wants to use the new AI model to screen data from those visits, identifying patients in need of follow-up for possible cognitive impairment.
In another effort, researchers running clinical trials on potential dementia treatments and diagnostic markers hope to use the AI tool to screen patients for enrollment. Often, clinical trial organizers find that 80% or more of the clinically eligible patients they recruit don’t meet the specific enrollment criteria for the trial. “Pharma companies are losing hundreds of thousands of dollars just based on patients [being excluded] during the screening process,” says Kolachalama. Augmenting the screening process with AI could reduce that cost significantly, and help direct patients toward trials for which they’re best suited.
Xue, C., Kowshik, S. S., Lteif, D., Puducheri, S., Jasodanand, V. H., Zhou, O. T., Walia, A. S., Guney, O. B., Zhang, J. D., Poésy, S., Kaliaev, A., Andreu-Arasa, V. C., Dwyer, B. C., Farris, C. W., Hao, H., Kedar, S., Mian, A. Z., Murman, D. L., O'Shea, S. A., ... Kolachalama, V. B. (2024). AI-based differential diagnosis of dementia etiologies on multimodal data. Nature Medicine, 30(10), 2977-2989. https://doi.org/10.1038/s41591-024-03118-z