Deep Learning-based Classification of Mild Cognitive Impairment Using 1D and 2D Features of Audio Signals

We proposed the time-frequency deep neural networks (TF-DNN) to detect mild cognitive impairment (MCI) based on audio data. The collected audio data were used to build the TF-DNN for the picture description of 73 MCI patients and 31 healthy control (HC) or normal test subjects who were clinically diagnosed. For the features used in the TF-DNN, we applied time-domain and frequency-domain processing, where one-dimensional (1D) and two-dimensional (2D) features were extracted. The 1D audio features included the zero-crossing rate (ZCR), the spectral centroid (SC), and the spectral spread (SS), while the 2D feature included the audio spectrogram. Te performance analysis result showed an accuracy of 90% for the MCI classification using the TF-DNN.