PAT-Net: A Method for Alzheimer's Detection Using Long-Duration Speech

This study presents an automated method based on temporal algorithms for detecting Alzheimer's Disease (AD) from long speech samples. Addressing the limitations of traditional speech detection methods in feature processing, we utilized Mel-Frequency Cepstral Coefficients (MFCCs) as key features and fed them into our innovative PAM-Net model. The PAM-Net model integrates TimesNet with Pyramid Attention Mechanism (PAM) and effectively extracts and processes temporal features through a Multi-Scale Module (CSCM). Additionally, based on an analysis of the characteristics of MFCC-based handcrafted features, we updated the activation functions to better suit the model. Experimental results on the NCMMSC2021 AD and ADRESS-M datasets demonstrate that the PAM-Net model achieved prediction accuracies of 89.5% and 76.1%, respectively, surpassing the baseline models by 9.7% and 2.2%. Furthermore, the model also exhibited a 73.1% accuracy on the Ivanova Spanish-language dataset, demonstrating robust cross-lingual performance. These findings not only highlight the potential of the PAM-Net model in speech detection but also provide a new technological avenue for the early diagnosis of Alzheimer's Disease.