Integrating acoustic, prosodic, and phonological features for automatic Alzheimer's detection

Early and accurate diagnosis of Alzheimer's disease (AD) is critical for effective intervention. While prior research has explored speech-based biomarkers for AD, the role of acoustic, prosodic, and phonological features, individually and in combination, remains underexplored.

This study systematically investigates a broad set of continuous acoustic and prosodic features alongside discrete phonological features for AD detection. Experiments were conducted using the publicly available ADReSS Challenge dataset to ensure comparability with prior work. We evaluated feature importance using four feature ranking methods and assessed classification performance using seven machine learning models. Both individual feature groups and their combinations were analyzed.

Acoustic features demonstrated strong predictive performance, and the best results were obtained by combining acoustic, prosodic, and phonological features. The optimal configuration, using ensemble learning, achieved an F1-score of 0.89. This performance surpasses previous speech-only approaches on the ADReSS dataset and approaches the best overall reported result. Additional validation on the Delaware corpus (focused on mild cognitive impairment) and a longitudinal speech sample from President Ronald Reagan yielded positive results.

These findings indicate that different layers of speech representation provide complementary information for detecting cognitive decline. The results suggest that speech characteristics beyond lexical content are highly informative for AD detection. This highlights the potential of multi-layered, non-invasive, and lexicon-independent speech analysis as a tool for early diagnosi sof Alzheimer's disease.