Early differentiation between Alzheimer's disease (AD) and frontotemporal lobar degeneration (FTLD) is a prerequisite for secondary prevention and targeted trial enrollment, yet remains challenging at disease onset. We investigated whether automated speech analysis could serve as a digital biomarker for early etiological stratification across clinically heterogeneous presentations.

In this prospective biomarker-confirmed prognostic study, 172 participants (108 patients with biomarker-confirmed AD or FTLD and 64 controls) completed a standardized speech protocol at initial clinical assessment. Acoustic, temporal, and phonatory features were automatically extracted. Machine learning models and a stacking ensemble were trained using stratified, repeated 5-fold cross-validation to discriminate between AD and FTLD pathology, with exploratory analysis extending to atypical and rare phenotypes crossed with physiopathology, including primary progressive aphasia (PPA) variants.

Speech-based models achieved high sensitivity and specificity in distinguishing independently physiopathology (mean area under the curve (AUC) = 0.986) and crossed phenotype and physiopathological diagnostic association (mean AUC=0.966).The ensemble identified 82% of cases with clinicopathological discordance. Interpretability analyses revealed distinct speech signatures: AD was associated with global speech slowing and phonatory instability, while FTLD was characterized by reduced verbal output and acoustic hypo-expressivity.

Automated speech analysis provides a promising non-invasive digital biomarker for the early etiological stratification of AD and FTLD, including atypical phenotypes, with high accuracy in a monocentric biomarker-confirmed cohort. These findings support the feasibility of speech-based etiological stratification and its potential to complement existing biomarker frameworks, particularly in cases of clinicopathological discordance. External validation is required before clinical deployment can be considered.