A Scoping Review of Foundation Models in Dementia Research

Dementia is a profound challenge for families and health systems alike, a condition that slowly erodes memory and independence. As the global population ages, the number of people living with this condition is rising, creating an urgent need for better ways to detect it early. Early detection is critical because new treatments are becoming available that work best when started at the very beginning of the disease, before significant damage occurs. To find these early signs, scientists are turning to artificial intelligence, specifically a powerful new type of computer program known as a foundation model. Think of these models as vast libraries of knowledge that have read millions of books, listened to countless conversations, and studied diverse images, allowing them to understand patterns in human language and behavior that a standard computer program might miss. Researchers hope these tools can listen to a patient's speech, read their medical notes, or analyze brain scans to spot the subtle fingerprints of dementia long before a doctor might notice them.

A team of researchers from universities in the United Kingdom recently set out to map the current landscape of this technology. They wanted to know exactly how these advanced AI systems are being used to study dementia, how well they actually work, and what gaps remain before they can be trusted in a hospital. To answer this, they conducted a wide-ranging review of scientific papers published between January 2022 and August 2026. They looked for studies that applied these foundation models to tasks like detecting the disease, diagnosing it, predicting how it might progress, or estimating a person's risk of developing it. After sifting through hundreds of records, they settled on fifty high-quality studies to analyze in depth, examining the methods used, the data fed into the computers, and the results achieved.

The researchers found that the field is moving fast but remains in its early stages. Most of the studies focused on two main tasks: detecting the presence of dementia or diagnosing it after symptoms appear. The way these models were built varied significantly. Some studies used models that were already trained on massive amounts of general data and then carefully adjusted, or "fine-tuned," for dementia tasks. Others used commercial AI systems, the kind available to the public, by simply asking them questions or giving them data without changing their internal settings. A smaller group of studies built models from scratch, training them entirely on new data. The results showed a clear trade-off between the size of the model and its reliability. The massive commercial systems produced the highest scores in some cases, but their performance was wildly inconsistent; in some instances, they performed no better than random guessing. In contrast, the models that were fine-tuned for the specific task of dementia research delivered the most consistent and dependable results, with the majority of studies showing strong performance. The models built from scratch generally performed moderately, often failing to outperform simpler, older types of computer programs.

Despite these promising technical results, the review uncovered serious limitations that prevent these tools from being used in real clinics today. A major issue was the lack of diversity in the data. The vast majority of the studies relied on data from the United States, and the participants were overwhelmingly white. Very few studies reported the race or ethnicity of the people involved, and many did not even report their ages. When gender was reported, the results were rarely broken down to see if the AI worked equally well for men and women. This lack of diversity means we do not know if these tools will work for people of different backgrounds, languages, or ages. Furthermore, the data used was often narrow, relying heavily on speech recordings or single types of medical records, rather than combining multiple sources like brain scans, genetic information, and lifestyle data.

Another critical finding was the weakness in how these models were tested. In medical research, a new tool must be proven to work on data it has never seen before to ensure it is not just memorizing the answers. Most of the studies reviewed only tested their models on the same data they used to train them, a practice that can lead to overly optimistic results. Only a small number of studies tested their models on completely separate groups of people from different hospitals or regions. Even fewer compared their results against the gold standard for dementia diagnosis, which is the examination of brain tissue after death. Without this rigorous testing, it is impossible to know if these AI systems can handle the complexity and variety of real-world patients. The researchers concluded that while the technology holds great promise, it is currently a proof of concept rather than a ready-made solution. To move from the laboratory to the clinic, future research must prioritize larger, more diverse datasets, rigorous testing on independent groups, and a clear understanding of how these tools perform across different populations. Until then, these powerful systems remain a work in progress, offering a glimpse of the future but not yet a replacement for human clinical judgment.