Over the past decade, artificial intelligence, AI, has been increasingly used for medical and neuroimaging research. While there have been great advances in detecting dementia with this technology, one of the remaining problems is the ability to apply predictions across multiple populations.
To correct this problem, scientists at The University of Texas Health Science Center at San Antonio created a mathematical equation that weighs information proportionally to have predictions be more generalizable. Importantly, the researchers found that differences in model performance persisted even after balancing the number of participants across groups and accounting for demographic and imaging-site differences, suggesting that unequal sample size alone does not explain the problem. Their study was published online June 26 in Nature Communications.
“Generalizing results to all populations was a clear problem we wanted to solve. There are two ways to solve this — collect as much data as possible from all groups or adjust the AI system by developing mathematical models that can leverage existing data where there is insufficient data for some groups,” said the study’s principal investigator, Mohamad Habes, PhD, director of the Neuroimaging Core and associate professor in the Department of Radiology at the Long School of Medicine and the Glenn Biggs Institute for Alzheimer’s and Neurodegenerative Diseases (Biggs Institute).
Building a better model
The study’s first author Ngoc-Huynh Ho, PhD, instructor in the Department of Radiology at the Long School of Medicine and the Biggs Institute, created a mathematical model that makes generalization possible for groups with limited data.
“We want to produce the proper probabilities for each group that we classify, but data is unbalanced for some populations. The goal was to guide the model to learn the probability of each class efficiently,” said Ho.
While AI is being used for disease diagnosis, researchers are also using explainable AI to understand what regions of the brain are contributing to these diagnostic decisions. Explainable AI is a method in which people can see and understand how a model comes up with a certain answer.
Research has uncovered differences in where and how the brain is affected in patients with dementia in different populations. Some of these differences may be due to co-morbid conditions, biology or the nature of the disease.
For example, some populations have higher vascular burden and that may contribute to differences in patterns seen in dementia.
These variations may explain why AI predictions are not always generalizable.
Making reliable predictions
“The predictions are highlighting brain regions that are important, but those regions might not be contributing to the model in the same way across different populations. Because we have less data for these populations, the model we developed gives higher weights to these regions for these populations where we have a limited number of data points,” said Habes.
AI predictions must be carefully evaluated and validated across multiple populations to make sure the model is reliable and efficient in diagnosis. Habes said before these models can be used in the clinical setting, they need to be validated not just within a particular group used in training data, but across as many populations as possible.
“This highlights the need for robust AI models that are objective. We are bringing a solution for when the number of data points is limited,” Habes said.
Habes’ lab is continuing the research by validating predictions across other groups. They are currently focused on United States populations but hope that the work can be scaled in the future to be relevant to populations all over the world.
“To be truly useful, these models must be perfect in accuracy and generalizability. This is the dimension Dr. Ho brought in by optimizing the algorithms,” Habes said.
While this study focused on dementia, Habes said the same principle can be applied to other neuropsychiatric or neurodegenerative disorders and even domains outside the brain. His lab aims to extend the research further to detect subtypes of different dementias and see how they vary in terms of patterns and detection accuracy across different populations.
Ho, NH., Charisis, S., Honnorat, N. et al. Advancing fair and explainable machine learning for neuroimaging dementia pattern classification in multi-racial and multi-ethnic populations. Nature Communications 17, 8026 (2026). https://doi.org/10.1038/s41467-026-74515-w