Presentation Information
[1Yin-A-27]Detection of COVID-19 from cough sounds using the COUGHVID dataset and examination of the explainability of deep learning models
〇Keita Takahashi1, Saki Katayama1, Ryoichi Chatani1, Yukiko Nagao1, Takuya Yoshimoto1 (1. Chugai Pharmaceutical Co., Ltd.)
Keywords:
Healthcare,Audio,XAI
In clinical practice, audio data is promising for respiratory disease screening due to its non-invasive collection and low
patient burden, and for early detection and remote diagnosis. As AI is increasingly considered for clinical diagnosis, its
explainability is important for reliability. In this study, we built a COVID-19 cough sound classification model using the
COUGHVID dataset and aimed to improve explainability by visualizing the prediction process. Audio data was converted to
MFCC features and classified using CNN, LSTM, and RNN models, evaluated by five-fold cross-validation. The RNN model
showed the highest accuracy. We segmented the MFCC features by time and coefficient and applied LIME to visualize their
contributions to the model predictions. Low-frequency bands contributed less, while mid-frequency bands tended to indicate
COVID-19 negative cases. This visualization helps explain the model predictions. Challenges include improving preprocessing
and modeling for better classification and extending applicability to other respiratory diseases.
patient burden, and for early detection and remote diagnosis. As AI is increasingly considered for clinical diagnosis, its
explainability is important for reliability. In this study, we built a COVID-19 cough sound classification model using the
COUGHVID dataset and aimed to improve explainability by visualizing the prediction process. Audio data was converted to
MFCC features and classified using CNN, LSTM, and RNN models, evaluated by five-fold cross-validation. The RNN model
showed the highest accuracy. We segmented the MFCC features by time and coefficient and applied LIME to visualize their
contributions to the model predictions. Low-frequency bands contributed less, while mid-frequency bands tended to indicate
COVID-19 negative cases. This visualization helps explain the model predictions. Challenges include improving preprocessing
and modeling for better classification and extending applicability to other respiratory diseases.