The Design of a Cough Disease Classification System Using Cube-Transformer

Cough is a common symptom of respiratory diseases. Hence, making a correct diagnosis of the respiratory diseases through cough classification is an important issue. The current methods for diagnosing COVID-19 include rapid screening and PCR testing. However, the cost of the above approaches is high and require the contact with patients and there is a risk of infection. Most patients diagnosed with COVID-19 have cough symptoms. Hence, diagnosis based on cough sounds is a cheaper and safer approach. In order to improve the quality of training data, the data conversion method of MFCCs is used in the preprocessing of the data. The target data is extracted from the sound data and then projected into the image data, and the sound is analyzed for the diagnosis of the respiratory diseases. Regarding the AI model for calculating the correlation between different data, the self-attention operation mechanism in Transformer can calculate the degree of the correlation between sequence nodes. Therefore, the Transformer model originally used in the translation field is used as the basis for improvement. Based on the concept of hypercube, a self-attention architecture with hypercube characteristics is proposed, which is named as Cube-Transformer. The Cube-Transformer model is mainly optimized and improved from the self-attention operation mechanism and the Star-Transformer model. The Cube-Transformer model can provide data with different self-attention computation scores on different planes and pay attention to the correlation of different data spaces. The experimental results show that the Cube-Transformer model can learn more feature aspects than the Star-Transformer model does and thus improve the identification accuracy by 1.5%.