Mandarin speech-based early detection of SCD: a feature-fusion residual network method

Introduction: Alzheimer's disease (AD) poses a global health challenge. Early intervention during the stage of subjective cognitive decline (SCD) - a potential window for delaying disease progression - is crucial. This study aims to assess an exploratory speech-based model for rapid SCD screening. Method: This study included 459 participants, comprising individuals with AD, mild cognitive impairment (MCI), SCD, and normal controls. We used Pic-Talk clips and Mandarin speech with residual network features for SCD screening. Results: In this cross-sectional study, our model achieved high performance with accuracy, recall, precision, F1, and area under the curve of 81.77 ± 2.78%, 80.53 ± 2.64%, 82.27 ± 2.38%, 81.39 ± 1.85%, and 82.85 ± 2.01%, respectively, outperforming other speech models. Discussion: This non-invasive exploratory approach to SCD assessment shows potential, revealing acoustic differences at the group level between SCD and other diagnostic groups. It is expected that the future integration of biomarkers will enhance the model's accuracy and expand its applicability.