Deciphering the Puzzle of Dementia

A diagnosis of dementia often disrupts the lives of patients and their families. AI-enabled algorithms can provide a more definitive diagnosis, and in some cases help detect curable forms of the condition.

By John Halamka, M.D., Diercks President, Mayo Clinic Platform and Paul Cerrato, MA, senior research analyst and communications specialist, Mayo Clinic Platform

The image of an older person staring into space or not recognizing loved ones continues to worry families that suspect a relative may be developing Alzheimer’s disease and some related form of dementia. And while families may suspect the worst, neurologists point out that there are numerous types of dementia, many of which can be successfully managed—and sometimes cured. 

Every three seconds, someone in the world develops dementia; the total cost of the condition is estimated to be over US$ 1.3 trillion worldwide. The list of possible causes of the condition is long, which makes it difficult for primary care physicians and nurses to tell them apart. Besides Alzheimer’s, clinicians also need to consider vascular dementia from cerebrovascular disease, dementia with Lewy bodies, Parkinson’s disease dementia, frontotemporal dementia, vitamin B12 deficiency, medication-induced dementia, and brain injury, among others. Several investigators have been attempting to create an AI-enhanced diagnostic tool to help differentiate between various causative agents. Boustani et al, for instance, have developed a machine learning based system that includes a Quick Dementia Rating System (QDRS) and a Passive Digital Marker (PDM) for use by primary care clinicians. They conducted a randomized clinical trial to test its value in patients 65 and older and found that combining the two assessment tools increased the odds of making an accurate diagnosis by 31%, when compared to usual clinic care.

David Jones, MD, a neurologist and the founding director of the Mayo Clinic Neurology Artificial Intelligence Program, has likewise used AI to differentiate various types of dementia. Using a tool called StateViewer, he and his colleagues have been able to identify several types of dementia in 88% of the cases they analyzed, according to their report in Neurology. To accomplish this result, they used a single, widely available brain scan, the fluorodeoxyglucose positron emission tomography (FDG-PET). The scan displays specific regions on a patient’s brain, and the unique patterns identified by the scans have been matched to specific types, or combinations, of dementia (See Figure). That in turn helps clinicians make a more accurate diagnosis and develop a personalized treatment plan that would not be possible otherwise.

Using the StateViewer, which relies on a modified k-nearest neighbor (K-NN) algorithm, Bernard et al “was able to detect the presence of 9 different neurodegenerative phenotypes with a sensitivity of 0.89 ± 0.03 and an area under the receiver (AUC) operating characteristic curve of 0.93 ± 0.02.”  In addition, they performed a radiologic reader study to compare their model to a standard protocol that’s used to evaluate FDG-PET scans. During that analysis, the researchers had radiologists differentiate between two disorders: posterior cortical atrophy and Lewy body dementia. The standard approach to diagnosis resulted in an AUC of 0.84. When StateViewer was used instead, AUC was 0.92. When the two approaches were combined, the researchers saw even better results: AUC 0.94.

Dementia may be a difficult diagnosis for patients and their families to cope with, but with the help of the latest AI algorithms, the diagnosis is becoming more precise, enabling clinicians and patients to chart a more personalized course of action.

------------------

Figure: The functional brain during neurodegenerative syndromes

Each colored node in the image represents an individual patient positioned within a low-dimensional functional state space derived from whole-brain metabolic network patterns. The closeness of the nodes to one another reflects similarity in large-scale brain network organization rather than clinical diagnosis. Nodes are color-coded by clinical syndrome, and broader regions of the space align with dominant cognitive–behavioral domains, including memory, semantics, language, executive function, behavior, visual processing, and action/motor planning.

Overall, the figure illustrates that neurodegenerative diseases are not discrete categories but represent structured, overlapping disruptions of large-scale generative brain networks, with clinical syndromes emerging from degeneration along functional gradients rather than from isolated regional damage within the brain. This framework provides a mapping of the degenerating mind that enables clinicians to chart trajectories toward precision diagnostic frameworks and to align current and emerging therapeutic strategies.

Pathologic / disease groupings:

• ADNC — Alzheimer’s Disease Neuropathologic Change
• FTLD — Frontotemporal Lobar Degeneration
• TDP-43 — TAR DNA-binding protein 43
• FTLD-TDP-43 — FTLD associated with TDP-43 pathology
• FTLD-Tau — FTLD associated with tau pathology
• Alpha-synuclein — Pathology dominated by α-synuclein protein aggregation
  

Clinical syndromes (node colors):

• Alzheimer-related / ADNC-associated
  • ad — Typical amnestic Alzheimer’s disease
  • lvppa — Logopenic Variant Primary Progressive Aphasia
  • dad — Dysexecutive Alzheimer’s disease
  • pca — Posterior Cortical Atrophy
• FTLD-TDP-43–associated
  • bvftd — Behavioral Variant Frontotemporal Dementia
  • lans — Limbic-predominant Age-related Neurodegenerative Syndrome
  • sd — Semantic Dementia
• FTLD-Tau–associated
  • nfppa — Nonfluent/Agrammatic Primary Progressive Aphasia
  • psp — Progressive Supranuclear Palsy
  • ppaos — Primary Progressive Apraxia of Speech
  • cbs — Corticobasal Syndrome
• Alpha-synuclein–associated
  • dlb — Dementia with Lewy Bodies
• Other
  • nph — Normal Pressure Hydrocephalus
  • cu — Cognitively Unimpaired