Single-Feature and Multi-Feature Fusion Audio Classification for Alzheimer's Disease Based on Convolutional Neural Network

EasyChair Preprint 6747

Abstract

Alzheimer's disease is a progressive brain disease, which worsens over time. Therefore, early screening of Alzheimer's disease is of great significance. Previously, the recognition of Alzheimer's disease relays on various acoustic features and speech transcription. This paper proposes an Alzheimer's disease audio classification method based on convolutional neural network, only uses the speaker's audio features, multi-feature fusion, combined with the ensemble learning method among models, and uses speech processing strategies such as multiple random sampling and speech endpoint detection. It can effectively distinguish the audio of patients with Alzheimer's disease, patients with mild cognitive impairment and normal people. The model proposed in this paper has achieved 84.87% (ranked fourth) and 83.78% (ranked third) accuracy in long audio track and short audio track respectively. In addition, this paper also explores the classification performance of Alzheimer's disease with different network structures and training methods. The source code can be accessed in https://github.com/lzl32947/NCMMSC2021_AD_Competition.

Keyphrases: Alzheimer detection, audio classification, machine learning, multi-feature fusion, speech processing

@booklet{EasyChair:6747,
  author    = {Zhilin Liu and Yanyu Yang and Zhao Yang and Kun Zhao and Wei Xi},
  title     = {Single-Feature and Multi-Feature Fusion Audio Classification for Alzheimer's Disease Based on Convolutional Neural Network},
  howpublished = {EasyChair Preprint 6747},
  year      = {EasyChair, 2021}}