Speech is emerging as a promising digital biomarker of brain health. However, most speech-based studies in MCI and dementia have focused on cross-sectional diagnostic classification rather than longitudinal prognosis. Although speech-derived features have recently been used to predict MCI-to-dementia progression within fixed follow-up windows, their value for estimating time-to-dementia conversion within a survival-analysis framework remains insufficiently characterized.

To assess whether spontaneous speech-derived features serve as prognostic digital biomarkers for time-to-dementia conversion in individuals with MCI, and to benchmark their performance against established cognitive composite measures.

We conducted a longitudinal prognostic study in patients with mild cognitive impairment evaluated at the Ace Alzheimer Center Barcelona Memory Clinic from 2022 onward. At baseline, participants completed a brief spontaneous speech protocol from which multi-domain linguistic and acoustic features were automatically extracted. Patients underwent standard annual clinical follow-up to determine conversion to dementia. Baseline speech features were used to train survival models estimating conversion risk for all participants, and performance was evaluated using the C-index, 2-year time-dependent AUC, and IBS. Predictive accuracy was benchmarked against models trained on cognitive composite scores.

A total of 853 individuals diagnosed with mild cognitive impairment were included at baseline, of whom 187 converted to dementia during follow-up. Cognitive composite models showed the strongest prognostic performance, achieving a test-set C-index of 0.77, 2-year horizon AUC of 0.82, and IBS of 0.12. Among speech-derived feature sets, Semantic Verbal Fluency showed the best performance, with a test-set C-index of 0.74, 2-year AUC of 0.74, and IBS of 0.12.

Semantic Verbal Fluency-derived speech features demonstrated good prognostic performance for MCI-to-dementia conversion, although cognitive composites remained the strongest predictors. These findings suggest that a brief, automatically processed verbal-fluency task may provide a scalable and low- burden complement to standard cognitive assessment for dementia risk stratification, particularly in settings where repeated or remote monitoring is desirable. External validation and longer follow-up are needed before clinical implementation.