Scholarlink

<p>Alzheimer’s disease (AD) is a debilitating neurodegenerative disorder that affects millions worldwide. Early and accurate diagnosis is crucial for timely intervention and management, as it can significantly improve patient outcomes and quality of life. This study explores a novel deep learning approach for AD detection using speech analysis, a promising non-invasive technique. We analysed the performance of various feature types, including magnitude spectrograms, raw phase, modulated group delay (MGD), and instantaneous frequency (IF) phase, on custom Convolutional Neural Network (CNN) and pre-trained ResNet50 models. The findings demonstrate that the magnitude spectrogram achieved the highest accuracy among single-input models. However, a dual-input model that combines magnitude and instantaneous frequency information from pre-trained ResNet50 models achieved superior performance, reaching an accuracy of 90%, precision of 95%, recall of 90%, and F1-score of 92.4%. This suggests that integrating magnitude and phase features effectively captures the nuances of speech patterns, leading to improved classification for AD detection. Moreover, the proposed approach outperforms previous state-of-the-art methods, highlighting its potential as a valuable tool for early AD diagnosis.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Speech signal’s phase information based Alzheimer’s disease detection using deep learning

  • 期刊论文

  • 发表日期：

    2025年5月14日

查看全文
  • Mohit Kumar,
  • Sushant,
  • Arun Kumar Yadav

摘要

Alzheimer’s disease (AD) is a debilitating neurodegenerative disorder that affects millions worldwide. Early and accurate diagnosis is crucial for timely intervention and management, as it can significantly improve patient outcomes and quality of life. This study explores a novel deep learning approach for AD detection using speech analysis, a promising non-invasive technique. We analysed the performance of various feature types, including magnitude spectrograms, raw phase, modulated group delay (MGD), and instantaneous frequency (IF) phase, on custom Convolutional Neural Network (CNN) and pre-trained ResNet50 models. The findings demonstrate that the magnitude spectrogram achieved the highest accuracy among single-input models. However, a dual-input model that combines magnitude and instantaneous frequency information from pre-trained ResNet50 models achieved superior performance, reaching an accuracy of 90%, precision of 95%, recall of 90%, and F1-score of 92.4%. This suggests that integrating magnitude and phase features effectively captures the nuances of speech patterns, leading to improved classification for AD detection. Moreover, the proposed approach outperforms previous state-of-the-art methods, highlighting its potential as a valuable tool for early AD diagnosis.