Análise de interpretabilidade em modelo profundo para detecção de insuficiência respiratória: um estudo de caso para a COVID­-19

Abstract

COVID-19 is a disease that has affected the whole world, being declared a pandemic by the World Health Organization. All methods created to detect this disease are costly and human, which is available in specific conditions so that the diagnosis is correct. Thus, in this work, the state of the art in detecting respiratory failure due to COVID-19 by ANNs was analyzed to verify biases and provide explanations about the result of the classification of this ANN. For this, an ablation test was performed, including new information as input to the model, the Grad-CAM algorithm was used to highlight which parts of the data fed to the ANN are more important. Also, audios of the product between an original audio and the Grad-CAM heat map were synthesized to allow a sound analysis of the results. In addition, a PANN was trained and also Mixup and SpecAugment techniques were used on training of SpiraNet to overcome the state of art. The results of the ablation test showed a great importance of the fundamental frequency of the voice and the melspectrogram. The synthesized audios showed that stressed syllables and prolonged words are important for the classification of patients in this work. The data augmentation experiment did not obtain significant results. And finally, the state of the art was surpassed with the trained PANN obtaining an accuracy of 94,44%.