Machine Learning Reveals Cracks in Dementia Diagnostic Boundaries

When a patient walks into a memory clinic complaining of personality changes or failing language, clinicians face one of neurology’s most stubborn puzzles: which dementia is this? A new study from the German FTLD Consortium suggests the answer may be harder to find than the official diagnostic rulebooks imply. By applying an unsupervised machine learning algorithm to neuropsychological test data from nearly 800 participants, researchers led by Marie Söntgerath of the Max Planck Institute for Human Cognitive and Brain Sciences and Matthias Schroeter of the University of Leipzig have shown that the cognitive fingerprints of frontotemporal lobar degeneration, or FTLD, do not always align with the clinical categories doctors use every day. The work, published in Frontiers in Dementia, is the first to jointly cluster patients with behavioral variant frontotemporal dementia, all three variants of primary progressive aphasia, amnestic Alzheimer’s disease, and healthy controls using neuropsychological data alone.

FTLD is the second most common cause of early-onset dementia after amnestic Alzheimer’s disease, yet diagnosing it can take more than three years, particularly in young-onset cases. The behavioral variant, known as bvFTD, erodes empathy, unleashes apathy and disinhibition, and often leads patients to be misdiagnosed with primary psychiatric disorders such as major depression. The language-predominant syndromes, collectively called primary progressive aphasia, come in three flavors: the semantic variant, which dissolves word meaning; the non-fluent variant, which makes speech effortful and telegraphic; and the logopenic variant, widely considered an atypical form of Alzheimer’s disease because of its underlying pathology. Each variant is supposed to carry a distinct neuropsychological signature, but in practice the boundaries blur.

To test whether those signatures are real, the team turned to k-means clustering, an algorithm that groups data points by similarity without ever being told the diagnostic labels. The method works by iteratively assigning each participant to the nearest cluster center in a multidimensional space defined by dozens of test scores, from memory recall and naming tasks to processing speed and companion-rated behavior questionnaires. Because the algorithm is unsupervised, any correspondence between the resulting clusters and the clinicians’ diagnoses is a genuine, data-driven finding rather than a foregone conclusion. The researchers ran the analysis repeatedly, varying the number of clusters from two to nine and systematically changing which diagnostic groups were included, then measured agreement with the clinical labels using a statistic called adjusted mutual information.

The dataset came from 13 university clinics and research centers across Germany. Of 1,036 original participants, 782 remained after exclusions, and after handling missing values through multiple imputation by chained equations, 469 participants entered the analysis without questionnaires and 462 entered the analysis with them. Diagnoses had been made by experienced neurologists or psychiatrists using established international criteria, supported in the majority of cases by brain imaging and cerebrospinal fluid markers. Only a small fraction, around six to seven percent, had neuropathological confirmation, reflecting the reality that definitive diagnosis of these diseases is still only possible after death.

The results were strikingly consistent across analyses. Healthy controls, patients with semantic variant PPA, and patients with bvFTD each formed comparatively homogeneous clusters that separated from other groups as the number of clusters increased. Healthy controls reached maximum homogeneity at nine clusters, with 80 percent purity when questionnaire scores were included. Semantic variant patients clustered apart from everyone else, and so did most bvFTD patients, though a subset of bvFTD patients drifted toward the healthy controls when behavioral questionnaires were left out of the analysis. That drift proved telling: when the Frontal Systems Behavior Scale and the Apathy Evaluation Scale, both filled out by companions, were added back in, bvFTD separated cleanly at lower cluster numbers, confirming that informant-rated behavior, not cognitive testing, is what truly distinguishes these patients.

The most provocative finding concerned the two non-fluent aphasia variants. Despite being defined as separate syndromes in the current diagnostic criteria, patients with non-fluent variant PPA and logopenic variant PPA consistently clustered together, indistinguishable by the cognitive and language tests applied. This mirrors recent neuroimaging work by the same consortium, in which machine learning classifiers identified semantic variant PPA with high accuracy but failed to separate the two non-fluent variants, operating almost at chance level. The authors suggest that either the neuropsychological tests used lack the sensitivity to detect the phonological working memory and speech apraxia deficits that theoretically distinguish the two syndromes, or that fluid biomarkers will be needed to draw that line reliably.

Amnestic Alzheimer’s disease posed its own puzzle. Rather than forming a distinct cluster, AD patients mixed with all other patient groups to a similar extent, separating cleanly only from healthy controls. When the analysis was restricted to just bvFTD and AD, maximum homogeneity for an AD cluster reached 85 percent with questionnaires included, but the separation remained incomplete. The neuropsychological profiles revealed why: memory impairment in bvFTD can be as pronounced as in Alzheimer’s disease, and the diagnostic criteria’s claim of relatively spared memory in bvFTD holds only in a relative sense. Alzheimer’s patients showed greater memory impairment relative to their overall disease severity, while bvFTD patients showed either globally preserved function or globally severe impairment, a pattern the researchers captured using the frontotemporal lobar degeneration clinical dementia rating.

Beneath the cluster boundaries, the profiles told a coherent story. Semantic variant patients scored worst on naming and semantic tasks, and showed a characteristic drop from repetition to comprehension on the repeat and point task, a signature consistent with their loss of word meaning. The non-fluent variants instead showed impairment in processing speed, short-term memory, and working memory, performing worse on tasks like the Trail Making Test, digit span, and the Stroop test. The authors note an open question: many of these so-called non-language tests still rely on language processing, so the observed deficits could be side effects of underlying linguistic breakdown rather than independent cognitive decline. Either way, the finding that cognitive involvement in PPA extends beyond the language domain, even at early disease stages, challenges the notion of circumscribed syndromes.

The study also delivered some negative results with real consequences. Standard executive function tests and a social cognition measure, the Reading the Mind in the Eyes Test, failed to separate bvFTD from other groups, despite previous literature suggesting they should. The authors attribute this partly to the single, semantically demanding social cognition test used and its high rate of missing data, and they call for alternative instruments such as the Social Cognition and Emotional Assessment or tasks observing spontaneous social behavior. They also acknowledge the inherent limitations of k-means clustering, which favors spherical, similarly sized clusters and may shape results in ways other algorithms would not.

Perhaps the deepest insight is that much of the apparent heterogeneity within diagnostic groups traces back to disease severity rather than genuine subtype differences. At low cluster numbers, severity drove the separation; only at higher cluster numbers did disease-specific patterns emerge. The authors argue that differential diagnosis may benefit from stage-specific comparisons, and they point toward tools like subtype and stage inference as a path forward. For a field entering the era of disease-modifying drugs, where accurate diagnosis is the prerequisite for the right treatment reaching the right patient, the message is clear: the cognitive boundaries between these dementias are fuzzier than the textbooks suggest, and the criteria that define them may need refinement to match what the data actually show.

Subject of Research: Data-driven clustering of neuropsychological profiles across frontotemporal lobar degeneration and Alzheimer's disease syndromes

Article Title: Evaluating the correspondence of neuropsychological profiles with clinical diagnostic criteria for frontotemporal lobar degeneration and Alzheimer's disease: insights from k-means clustering

Article References: Evaluating the correspondence of neuropsychological profiles with clinical diagnostic criteria for frontotemporal lobar degeneration and Alzheimer's disease: insights from k-means clustering. (n.d.). https://doi.org/10.3389/frdem.2026.1757680

Image Credits: AI Generated

DOI: 10.3389/frdem.2026.1757680

Keywords: frontotemporal lobar degeneration, behavioral variant frontotemporal dementia, primary progressive aphasia, Alzheimer's disease, k-means clustering, machine learning, neuropsychological testing, differential diagnosis, semantic variant PPA, logopenic variant PPA, non-fluent variant PPA, diagnostic criteria