Zampeta-Sofia Alexopoulou, Gonzalo Pérez, Hali Lindsay, Johannes Tröger, Nicklas Linz, Philippe Robert, Loreto Olavarría, Patricia Lillo, Daniela Thumala, Cecilia Okuma, Mauricio Cerda, Fernando Henríquez, Andrea Slachevsky, Valeria Manera, Adolfo M. García, Alexandra König
*Poster presented at AAIC 2026
Background: Neuropsychiatric symptoms (NPSs) are important prognostic factors in
Alzheimer’s disease (AD). Speech analysis is a promising approach for capturing NPSs in
early-AD stages, including mild cognitive impairment (MCI). However, existing evidence
largely derives from single-language cohorts, limiting cross-linguistic generalizability. Here,
we address this gap by examining speech-derived markers of NPSs across two linguistically
and culturally distinct cohorts.
Methods: We analyzed data from a French-speaking cohort from France (39 with MCI, 30
healthy controls(HC)) and a Spanish-speaking cohort from Chile (106 with MCI, 84 HC).
Participants completed two standardized free-speech tasks, recalling one positive and one
negative autobiographical story. Audio-recordings were preprocessed using TELL’s v.2.0
pipeline. Acoustic (i.e. spectral, frequency) and linguistic (i.e. syntax, sentiment) features
were extracted using ki:elements’ AI-driven pipeline. NPSs severity was assessed with the
Neuropsychiatric Inventory (depression, anxiety, apathy). Within-tasks/diagnostic groups,
feature-values were compared between languages using Wilcoxon tests (multiple-
comparisons corrected). Elastic net models (α = 0.3) predicted NPSs severity, adjusting for
diagnosis, age, sex, education, and MMSE, with regularization selected via cross-validation.
Feature-selection robustness was assessed via bootstrapping (1.000 iterations; 80%
subsamples), computing selection frequencies for non-zero coefficients.
Results: 105 speech features were analyzed with several being significantly different between
French and Spanish speakers. Accordingly, models were fitted separately by language/task.
Despite overall low NPSs burden, models identified speech-based predictors of depression,
while fewer predictors emerged for apathy/anxiety. Although individual selected features
differed between languages, certain depression-related predictors converged across cohorts
within shared feature-domains, including lexical-syntactic (conjunction use, verb-phrase
composition), temporal (pauses/utterance durations) and spectral dimensions (Figs. 1-3).
Model performance was highest for depression (R² up to 0.68 for French; up to 0.19 for
Spanish). Speech features showed stronger associations with NPSs than
cognitive/demographic covariates, with sex most consistently retained.
Conclusion: Despite low NPSs severity, similar types of speech-features (rather than identical
individual features) were associated with depression across two languages. These findings
support free-speech features as potentially cross-linguistic markers for NPSs assessment in
MCI, inviting validation in additional languages and clinically enriched samples.