Background

Automated speech-and-language analysis (ASLA) is emerging as a scalable approach to capture markers of Alzheimer’s disease (AD) and behavioral variant frontotemporal dementia (bvFTD). Prior work shows these markers aid differential diagnosis, preclinical detection, symptom-severity prediction, and conversion risk in unimpaired individuals. Yet few studies test marker robustness to key speech and language modulators such as sex, age, education, dialect, languages, and social determinants of health (SDH). Here we evaluated whether well-established ASLA features withstand socio-cultural heterogeneity in a large, diverse Latin American cohort.

Methods

We evaluated whether well-established ASLA features discriminate AD and bvFTD from healthy controls (HC) and predict cognitive outcomes in a large Latin American cohort. The sample included up to 1,689 Spanish-speaking participants from Argentina, Chile, Mexico, Peru, and Colombia, comprising 670 AD patients, 106 bvFTD patients, and 913 HC. Participants completed two phonemic fluency, two semantic fluency, story retelling, and video retelling tasks. ASLA features captured speech timing, pitch, verbosity, concreteness, emotional content, granularity, lexical similarity, semantic similarity, and text topology. Machine-learning models included residualization for age, sex, education, dialect, number of languages, and eight social determinants of health. Performance is reported in the hold-out dataset.

Results

Residualized ASLA features showed strong hold-out discrimination for both AD-vs-HC and bvFTD-vs-HC. For AD-vs-HC, ASLA achieved AUC = 0.938 [0.926–0.950], nominally exceeding cognitive-domain models based on episodic memory, AUC = 0.924 [0.908–0.939]; language, AUC = 0.890 [0.874–0.905]; executive functions, AUC = 0.874 [0.856–0.891]; and socioaffective measures, AUC = 0.757 [0.731–0.780]. For bvFTD-vs-HC, ASLA achieved AUC = 0.860 [0.818–0.901], also exceeding cognitive-domain models based on language, AUC = 0.848 [0.804–0.887]; episodic memory, AUC = 0.813 [0.760–0.861]; executive functions, AUC = 0.804 [0.758–0.849]; and socioaffective measures, AUC = 0.721 [0.669–0.775].

ASLA performance remained high across the age and sex strata shown in the poster. For AD-vs-HC, AUCs were 0.913 [0.888–0.937] in participants aged ≤68 and 0.940 [0.922–0.955] in those aged >68; and 0.936 [0.922–0.949] in males and 0.945 [0.922–0.963] in females. For bvFTD-vs-HC, AUCs were 0.825 [0.755–0.889] in participants aged ≤68 and 0.899 [0.846–0.941] in those aged >68; and 0.899 [0.852–0.939] in males and 0.803 [0.724–0.874] in females.

ASLA also tracked cognitive-domain composites. In AD, correlations were significant for language, R = 0.795 [0.742–0.837]; episodic memory, R = 0.692 [0.635–0.742]; and executive functions, R = 0.583 [0.506–0.653]. In bvFTD, correlations were significant with values: R = 0.524 [0.244–0.701], R = 0.625 [0.442–0.777], and R = 0.548 [0.346–0.696], respectively. Socioaffective correlations were not shown because they were not significant.

Conclusions

These findings suggest that ASLA markers can support socioculturally robust detection and phenotyping of AD and bvFTD in Latin American populations. Automated speech-based tools may facilitate scalable screening and monitoring in low-resource or remote settings while reducing examiner-dependent variability.