Abstract

Verbal fluency tasks are ubiquitous in mild cognitive impairment (MCI) screenings. Yet, their assessment is traditionally limited to valid response counts. This subjective approach constrains analysis to univariate methods and overlooks which semantic memory dimensions are affected, introducing human bias while limiting informativeness. We tackled these gaps with a novel automated framework. Ninety-six participants (53 with MCI, 43 cognitively unimpaired individuals) performed phonemic and semantic fluency tasks alongside standard cognitive tests. Word properties (e.g., frequency, granularity, length) and timing features (e.g., number of pauses) were (i) automatically extracted to discriminate between groups via machine learning classification, (ii) benchmarked against standard cognitive measures (Trail Making Test-A, Trail Making Test-B, episodic memory subscore from the Addenbrooke's Cognitive Examination, digit span, and Mini-mental State Examination), and (iii) used to predict brain patterns and plasma phosphorylated tau 217 (pTau217) concentration. Our approach yielded robust classification performance when using word properties and speech timing features combined (Area under the receiver operating characteristic curve [AUC] = 0.81, 95% confidence interval [CI] = [0.71, 0.89]), outperforming cognitive measures (AUC = 0.77, CI = [.68, 0.85]). Frequency, granularity, and semantic distance correlated with the gray matter volume of semantic-related regions commonly atrophied in MCI. No fluency feature was associated with functional connectivity patterns. Granularity was moderately associated with pTau217 levels. In sum, automated fluency analyses facilitate MCI detection, capturing fine-grained neurocognitive and biomarker patterns in the condition.