Speech based detection of preclinical Alzheimer's disease in free conversation

This study proposes the Hidden-Unit BERT (HuBERT) approach for predicting preclinical Alzheimer’s disease using free-conversation speech, aiming to replace high-cost detection methods such as positron emission tomography (PET). Previous studies mainly focused on the mild cognitive impairment stage, whereas this study targets the preclinical stage where cognitive functions have not yet significantly declined, demonstrating a certain level of innovation. We employ the HuBERT model (rinna/japanese-hubert-base), pre-trained for Japanese, to extract audio features from speech, and construct a classification model to distinguish between amyloid-β positive/tau positive (A + T+) and amyloid-β negative/tau negative (A–T–) populations diagnosed by PET. Compared with traditional methods using acoustic features such as eGeMAPS, the proposed method more effectively captures subtle information in speech reflecting cognitive changes. Experimental results show that it achieves 15% higher recall and 18% higher AUC than conventional methods. This study highlights the potential of daily conversational speech in early Alzheimer’s detection and suggests it could serve as a low-cost, non-invasive screening tool.