Classification of Diagnosis of Alzheimer’s Disease Based on Convolutional Layers of VGG16 Model using Speech Data

摘要

In this study, dementia and stages of dementia were classified using Mel-Spectrogram and VGG16 models of speech data. The human voice is highly indicative of dementia and can also be used to infer its degree of progression. Therefore, meaningful features extracted from speech data can be used as indicators to determine the presence of dementia and its progression. Speech data were used from a total of 292 subjects diagnosed with Alzheimer's disease (AD) / mild cognitive impairment (MCI) / subjective cognitive impairment (SCI). The speech data were converted into Mel-spectrograms and used. We used the convolutional layers of the VGG16 model as a feature extractor of the Mel-spectrograms. The Pearson correlation coefficient between the extracted features and the labels was obtained to select the features that are effective for classification. As a result of 5-fold CV using the selected features, SCI vs. others (MCI and AD) showed an average classification accuracy of 90%, and a maximum classification accuracy of 93%. In the MCI vs AD group, the average classification accuracy was 84%, and the maximum classification accuracy was 90%. These results show that the Mel-spectrograms of speech data can provide useful information for confirming AD, MCI, and SCI.