Background: Verbal fluency tasks detect semantic memory deficits in general, and particularly in people with semantic variant primary progressive aphasia (svPPA). However, standard word count scoring is effortful, coarse, and often inconsistent, limiting applicability, reliability, and cross-study comparability. Evidence from English-speaking cohorts suggests that these limitations are overcome by digital word property analysis. Yet, this approach has not been tested in Spanish-speaking Latinos with svPPA, an underserved, high-risk population with dire need of scalable testing tools. Here we tackle this gap.

Methods: Seventy-seven participants from the ReDLat consortium (36 svPPA patients, 41 demographically-matched controls) completed semantic (‘animals’) and phonemic (/p/) fluency tasks. We computed item-level word properties (frequency, phonemic length, phonological neighbors, concreteness, granularity, semantic and phonological variability), alongside whole-task audio features (pitch, voice quality, speech timing). Features were fed to a Logistic Regression classifier to discriminate between groups, with 5-fold nested cross-validation, recursive feature elimination and Bayesian hyperparameter optimization across ten seeds. Scores were bootstrapped to obtain 95% confidence intervals.

Results: For the semantic fluency task, word properties yielded robust classification [Logistic Regression: ROC-AUC = 0.91 (0.81-0.96)], surpassing audio features alone [0.77 (0.66-0.87)] and matching the combination of both features types [0.91 (0.82-0.95)]. Chance-level results were obtained upon shuffling labels [0.49 (0.39-0.57)]. Top features were granularity, number of phonemes, concreteness, and semantic variability. Conversely, analyses of phonemic fluency alone yielded near-chance results [Word properties = 0.57 (0.44-0.67); audio = 0.59 (0.49-0.70); combination = 0.61 (0.50-0.72)]. Combining both tasks did not improve performance [best ROC-AUC = 0.90 (0.81-0.95)].

Conclusions: Digital word property analyses of semantic fluency tasks seems to support svPPA detection in Spanish-speaking Latinos, arguably capturing degraded semantic memory search skills. These tools provide an objective, automated resource for scalable clinical testing in this underserved population.