The task is disarmingly ordinary. Sit in a room, think back to a story you have known since childhood, and tell it: the stepsisters, the slipper, the coach. In the longitudinal Northwestern University PPA Research Program, that retelling of Cinderella has been recorded for years from people whose language is being taken from them by a neurodegenerative disease, and from cognitively healthy volunteers who tell it fluently and go home. What a team led by Daniel Gutstein, with Michael Iorga, Jane Stocks, Tamar Gefen, M. Marsel Mesulam and Elena Barbieri, asked of those recordings, in Alzheimer's & Dementia on 20 March 2026, is a question no clinician can answer at the bedside: not which variant of primary progressive aphasia this is, but which protein is causing it.

That distinction is the whole point. PPA is a syndrome, not a disease: it is what happens when Alzheimer's pathology or a form of frontotemporal lobar degeneration lands first in the brain's language network. The clinical variant only probabilistically predicts the pathology underneath, and the authors say so directly, noting that agrammatic PPA in particular is neuropathologically heterogeneous, which limits variant labels as proxies for what is actually in the tissue. Their study drew on 15 cognitively healthy, age matched controls and 77 people diagnosed with PPA, 54 with pathology confirmed at autopsy and 23 with a pathological diagnosis inferred during life. Transcripts were run through open natural language processing toolkits (SpaCy with TextDescriptives, and Stanza) to yield classical linguistic features, and separately through transformer embeddings; ensembles of 100 small neural networks were trained, each ensemble predicting the class of a single held out participant by majority vote, with the 23 inferred cases sitting in every training set. The AD and FTLD-4R-tau groups did not differ significantly in age at onset, Clinical Dementia Rating, or Western Aphasia Battery aphasia quotient, so the models were not simply reading off severity. All models separated AD from FTLD-4R-tau, and controls from PPA.

Then comes the sentence that a careful reader should sit with. In the discussion, the authors report that classical features relating to fluency influenced classification within both pathologies: FTLD-4R-tau is typically characterised by slower, less fluent output, and PPA due to Alzheimer's pathology also shows reduced fluency compared with cognitively healthy speakers. A feature that is degraded on both sides of the line is a strange thing to hang a pathology call on, even when the model is right. The authors also concede that the absence of comorbid pathologies in the inferred cases cannot be confirmed until autopsy. This is a retrospective study in one deeply characterised research cohort, tested by holding out one participant at a time. Nobody has yet handed the model a new patient at a new centre and waited to see what the tissue or the biomarkers said.

It is not alone in the attempt. In npj Dementia on 1 June 2026, Cole Robertson, Daisy Hochberg, Megan Quimby, Phillip Wolff, Bradford Dickerson and Neguine Rezaii reported using a pre-trained language model (Distil-RoBERTa) on short connected speech samples from 71 people with PPA, with amyloid status established by cerebrospinal fluid, amyloid PET or autopsy; 51 per cent were amyloid positive. Across ten runs of tenfold cross validation the classifier reached a mean accuracy of 92 per cent, and the team used explainable AI methods to pull out the linguistic patterns the classifier was leaning on. Earlier work in the field, such as the 2024 Alzheimer's & Dementia paper from Katheryn Cho and colleagues using one minute speech samples from 179 participants, had already shown that speech features can sort Alzheimer's neuropathologic change from FTLD. Cross validation, though, is a measure of internal consistency, not of what happens in a new clinic with a new microphone.

And the microphone matters more than the marketing suggests. A medRxiv preprint posted on 24 February 2026, not yet peer reviewed, took picture descriptions from 109 people with PPA and 32 healthy controls and compared manual transcription, automated transcription with Whisper, and a semi-automated route with a quality control step. Error rates were lowest for controls, then the semantic, logopenic and nonfluent variants, and errors correlated with disease severity in the semantic and logopenic groups; the quality control step cut errors and improved the reliability of the derived linguistic features. Classification of nonfluent PPA against controls was the hardest task, yet the quality controlled automated features reached an area under the curve of 0.89, above the 0.81 from manually derived features. Read together, the three papers describe a plausible cheap triage test and, in the same breath, all the places it can quietly go wrong: the feature that is not pathology specific, the validation that never leaves its own cohort, the transcription layer that misreads the very disfluencies the model is trying to weigh.