New Dutch Language Screen Spots Primary Progressive Aphasia With Near-Perfect Accuracy

A fifteen-minute language test has shown remarkably good accuracy at spotting primary progressive aphasia, a neurodegenerative syndrome in which people progressively lose the ability to speak and understand words. In a new study published in the Journal of Neurology, researchers adapted the Mini-Linguistic State Examination into Dutch and validated it in 90 Dutch-speaking patients alongside 276 healthy volunteers, achieving an area under the receiver operating curve above 0.99 for separating patients from controls. That figure, effectively near-perfect discrimination in this cross-sectional validation, marks an important step toward closing a long-standing gap: most speech and language assessments for this condition were built for and tested in English speakers, leaving clinicians elsewhere without validated instruments.

Primary progressive aphasia is a clinical syndrome in which neurodegeneration targets speech and language abilities first, with deterioration that is progressive over time. Current consensus criteria divide the syndrome into three canonical variants. The nonfluent or agrammatic variant is dominated by effortful, distorted speech production and grammatical breakdown. The semantic variant attacks word meaning, leaving patients struggling with naming and semantic knowledge. The logopenic variant impairs phonological working memory so that word-finding and sentence repetition falter. A substantial fraction of patients, however, fit none of these cleanly and are labeled Not Otherwise Specified, a category the new study deliberately included.

Diagnosing these variants properly has long required specialist expertise, extensive neuropsychological batteries, brain imaging, and fluid biomarkers, resources that are not readily available everywhere. Most existing screening tools probe only a single language domain, such as naming or verbal fluency, and therefore miss the broader profile of deficits that distinguishes one variant from another. The Mini-Linguistic State Examination was designed in English to fill that niche: a compact battery whose eleven subtests sample five domains, namely motor speech, phonological structure, semantic knowledge, syntax, and verbal working memory. Until now it had been formally validated only in English, Spanish, and Greek, and no Dutch results had been published.

Adapting the test into Dutch was a linguistic exercise in its own right, not a simple translation. The team created two independent versions for the variants of Dutch spoken in the Netherlands and Flanders, then reconciled them into a single instrument. Stimuli were matched to the English original on word frequency, age of acquisition, syllable count, and lexical stress patterns, the latter a known marker of apraxia of speech. Where English words had no suitable Dutch equivalents, new pseudowords were generated with the multilingual tool Wuggy to preserve phonotactic structure. A few items changed outright: Dachshund became Dalmatian and Chrysanthemum became Orchid, because the Dutch word frequencies differed. Sentence comprehension items retained their subordinate clauses and passive constructions to keep syntactic complexity intact, though sentence repetition stimuli gained roughly one syllable on average, slightly raising the working memory load relative to the English version.

The patient cohort spanned four academic memory clinics in Belgium and the Netherlands and covered the full clinical spectrum. Twenty-three patients had the nonfluent variant, twenty-four the logopenic variant, thirty-two the semantic variant, and eleven were classified as Not Otherwise Specified, with a mean age of 68.3 years. All diagnoses were established a priori using clinical information, neuropsychological testing, structural MRI, and, where available, fluid biomarkers, without any input from the new test. Five of the 24 patients with the logopenic variant had lumbar puncture data available, and in all five cases it confirmed Alzheimer’s disease pathology. The control group, recruited through social media, posters, and local cohorts, ranged in age from 18 to 85 to build a normative dataset across the adult lifespan.

The results replicated the expected neurolinguistic signatures of each variant with striking fidelity. Patients with the semantic variant showed a distinct and selective collapse in semantic knowledge, scoring below every other group, while their motor speech remained relatively preserved. Those with the nonfluent variant showed widespread impairment dominated by motor speech deficits, scoring far lower than the other two canonical variants on that domain. Logopenic patients were hit hardest in working memory and phonological structure, lower than the semantic group in phonology and lower than the nonfluent group in working memory. The Not Otherwise Specified group showed deficits across every domain except motor speech, consistent with its broad definition. Syntax errors, notably, were frequent in all variants.

In raw numbers, patients’ total scores ranged from 30 to 96 out of 100, while healthy volunteers scored between 83 and 100. Domain-level areas under the curve ranged from 0.80 to 0.92, with motor speech best separating nonfluent patients from controls at 0.96, semantic knowledge best capturing the semantic variant at 0.99, and working memory strongest for the logopenic group at 0.98. A sensitivity analysis stratifying fifty-four patients by disease stage found excellent discrimination even at the prodromal stage, with an area under the curve of 0.958 for the earliest cases. Inter-rater reliability, assessed in 123 participants, was good to excellent across all scores, with intraclass correlation coefficients between 0.78 and 0.97.

Beyond classical statistics, the researchers reused the machine-learning framework of the original English study, training a random forest classifier on the five domain scores. Against a demographically matched subset of 25 controls, the model classified held-out cases with 74 percent overall accuracy and balanced accuracy of 0.75. A simpler, interpretable decision tree extracted from the forest reached a balanced accuracy of 0.83, correctly identifying every control, 85 percent of semantic variant cases, 80 percent of logopenic cases, and 67 percent of nonfluent cases. Misclassifications were informative: some nonfluent patients with unusually preserved motor speech scores were pulled toward the logopenic category, and several semantic variant patients with reduced syntax and working memory scores drifted into the logopenic profile. Adding the heterogeneous Not Otherwise Specified group dropped balanced accuracy to 0.66.

The team also quantified how demographics shape test performance in healthy people. Older age and male sex were associated with lower total scores, while more years of education predicted higher scores, mirroring patterns seen for general cognitive screens in Dutch cohorts. Domain scores behaved differently: motor speech declined with age but was unaffected by sex or education, whereas syntax and working memory correlated with both age and education. Total scores correlated modestly with the Montreal Cognitive Assessment and the Boston Naming Test, with semantic knowledge showing the strongest link to naming. The authors suggest that age and education adjustment may sharpen interpretation, and their published percentile tables by age, sex, and education provide clinicians with demographically matched reference points, a resource often missing for Dutch-adapted neuropsychological tests.

Cross-linguistic validation matters more than logistical convenience suggests, because growing evidence indicates that primary progressive aphasia does not present identically across languages. Italian-speaking patients with the nonfluent variant make more syntactic errors while English speakers show more distortions, and Chinese-speaking patients display error patterns shaped by lexical tone and logographic script. Against that backdrop, the Dutch results closely matched the English, Spanish, and Greek versions: the same canonical profiles emerged, though the decision tree performed somewhat below the English version’s 91 percent, which the authors attribute to Dutch’s richer morphology or to a broader range of disease severity in their cohort. The study’s own limits include heterogeneous disease stages, limited biomarker confirmation of underlying pathology, and a modest Not Otherwise Specified sample. The researchers position the test as a triage and referral instrument rather than a diagnostic endpoint, and propose it could eventually track language decline over time in clinics and trials, pending longitudinal validation.

Subject of Research: Validation of a Dutch adaptation of the Mini-Linguistic State Examination as a screening tool for primary progressive aphasia

Article Title: The Dutch Mini-Linguistic State Examination (MLSE-NL): findings in Primary Progressive Aphasia (PPA)

Article References: Coppieters, R., Janssen, N., Garrard, P., Hiemstra, F., Duits, A., De Coninck, M., De Keulenaer, S., Storey, I., Piai, V., Renssen, J., Ruis, C., Seelaar, H., van den Berg, E., Van Mossevelde, S., Bruffaerts, R., & Jiskoot, L. (2026). The Dutch Mini-Linguistic State Examination (MLSE-NL): findings in Primary Progressive Aphasia (PPA). Journal of Neurology, 273(11), Article 684. https://doi.org/10.1007/s00415-026-14203-3

Image Credits: AI Generated

DOI: 10.1007/s00415-026-14203-3

Keywords: primary progressive aphasia, MLSE-NL, language assessment, neurodegeneration, aphasia screening, neuropsychology, logopenic variant, semantic variant, nonfluent variant, Dutch validation, random forest classifier, Journal of Neurology