Automated linguistic and acoustic measures of verbal fluency as neurocognitive markers of mild cognitive impairment

Abstract Background Verbal fluency tasks are routinely employed in screening for mild cognitive impairment (MCI). Yet, traditional outcome measures focus on the number of valid responses, failing to reveal which specific semantic memory dimensions may be altered and limiting analyses to univariate methods. Building on recent findings on Alzheimer’s disease, we employed automated methods to establish which linguistic and speech timing features better discriminate MCI patients from healthy controls (HCs), including machine learning analyses, comparisons with standard neuropsychological tests, and brain‐behavior correlations. Method We recruited 106 native Spanish speakers (52 with MCI, 54 HCs), who completed phonemic and semantic fluency tasks as well as standard tests of attention (Trail Making Test‐A) and episodic memory (Free and Cued Selective Reminding Test). Responses in the fluency tasks were audio‐recorded and transcribed for automatic extraction of word properties (granularity, frequency, phonological neighborhood, word length, imageability, familiarity) and speech timing features (number of syllables, pauses, pause duration, phonation time, articulation rate, average syllable duration). These variables were compared between groups through a generalized linear model (GLM, with standard cognitive test scores as covariates) and fed into a binary classifier for subject‐level discrimination. Structural and functional brain measures were obtained from 63 participants and subjected to voxel‐based morphometry and seed‐to‐voxel resting‐state connectivity analyses. Correlations between behavioral and brain features were examined via multiple regressions. Result GLM analysis revealed significant group differences in specific word properties (granularity, frequency, word length, imageability), but not in speech timing features. Subject‐level classification was better when based on word properties (AUC = 0.72) than on speech timing features (AUC = 0.63), with maximal discrimination upon combining both dimensions (AUC = 0.78). MCI patients exhibited atrophy left temporal atrophy and altered connectivity between the right parahippocampus and fronto‐posterior cortical regions. Word frequency was negatively correlated with insular volume and granularity was positively correlated with the volume of fusiform and parahippocampal regions. Conclusion Automated analyses of word properties and timing features in verbal fluency tasks offer novel insights into cognitive decline, surpassing traditional tests in identifying MCI. Their widespread implementation could inaugurate a promising avenue to establish scalable markers of the condition.