An automated machine learning approach to language changes in Alzheimer’s disease and frontotemporal dementia across Latino and English-speaking populations

PROJECT SUMMARY Alzheimer's disease (AD) and frontotemporal dementia (FTD) are highly prevalent in Latinos, the largest and fastest-growing minority in the United States (US). Yet, due to financial and cultural inequities, this group is challenged to afford standard diagnostic and monitoring procedures. Also, research on Latinos lacks scalable,

culturally valid tests and it rarely examines whether potential markers are robust across socio-biological profiles. Such issues can be tackled with low-cost automated speech and language analyses (ASLA). Participants are asked to produce natural speech, generating multiple acoustic (sound wave) and linguistic (e.g., semantic) data

that can be digitally extracted and analyzed to identify diseases or predict neurocognitive disruptions. Yet, ASLA findings are minimal in Latinos. Also, most ASLA studies are small and very few ha differentiated between AD and FTD variants, compared ASLA with standard measures, accounted for socio-biological factors (e.g., sex,

race, brain profile, bilingualism) or tested for validity across languages and dialects. This project will develop a novel ASLA framework to jointly address such challenges. To capture socio-biological diversity and meet requisites for robust machine and deep learning analyses, we will leverage 2740 participants.

These encompass Spanish speakers from five Latin American countries (700 AD, 700 FTD, 800 controls), English speakers from the US (140 AD, 140 FTD, 160 controls), and US-based Latinos (30 AD, 30 FTD, 40 controls), including the main variants of each disease. This is possible due to a strategic partnership between

UCSF and the Consortium to Expand Dementia Research in Latin America, a multi-funded network bringing a fully harmonized environment and a large, growing cohort. The Global Brain Health Institute, a dementia training hub at UCSF, hosts expert clinicians in all sites. Speech and language data will be gleaned through our new

Toolkit to Examine Lifelike Language, a HIPPA-compliant app for speech collection, storage, and visualization, supported by a language battery and survey. Enrollees are characterized with demographic, clinical, cognitive, and social determinants of health measures, alongside MRI and fMRI. Our ASLA approach comprises top

predicted markers for each syndrome, added fine-grained features, and embedding features. Novel machine and deep learning algorithms for high-dimensional settings will be used to pursue three aims. In Aim 1, we will employ machine and deep learning to reveal the ASLA markers that best identify AD and FTD

syndromes; compare them with cognitive and imaging measures; and test them for generalizability from Spanish onto English (a typologically different language). In Aim 2, via linear regressions, we will use optimal ASLA markers to capture syndrome-specific patterns of cognitive dysfunction, brain atrophy, and connectivity. In Aim

3, using high-dimensional machine learning, we will test such markers for validity across diverse socio-biological profiles, dialects, and bilingual skills (null, low, high). We will forge an affordable, scalable approach to assist AD and FTD diagnosis in Latinos, at a time when disease-modifying therapies may emerge.