Deep Recurrent Neural Networks with Attention Mechanisms for Respiratory Anomaly Classification

Abstract

In recent years, a variety of deep learning techniques and methods have been adopted to provide AI solutions to issues within the medical field, with one specific area being audio-based classification of medical datasets. This research aims to create a novel deep learning architecture for this purpose, with a variety of different layer structures implemented for undertaking audio classification. Specifically, bidirectional Long Short-Term Memory (BiLSTM) and Gated Recurrent Units (GRU) networks in conjunction with an attention mechanism, are implemented in this research for chronic and non-chronic lung disease and COVID-19 diagnosis. We employ two audio datasets, i.e. the Respiratory Sound and the Coswara datasets, to evaluate the proposed model architectures pertaining to lung disease classification. The Respiratory Sound Database contains audio data with respect to lung conditions such as Chronic Obstructive Pulmonary Disease (COPD) and asthma, while the Coswara dataset contains coughing audio samples associated with COVID-19. After a comprehensive evaluation and experimentation process, as the most performant architecture, the proposed attention BiLSTM network (A-BiLSTM) achieves accuracy rates of 96.2% and 96.8% for the Respiratory Sound and the Coswara datasets, respectively. Our research indicates that the implementation of the BiLSTM and attention mechanism was effective in improving performance for undertaking audio classification with respect to various lung condition diagnoses.

A Deep Ensemble Neural Network with Attention Mechanisms for Lung Abnormality Classification Using Audio Inputs

“…Such three-class lung condition detection is comparatively less challenging in comparison with the recognition of six different lung abnormalities. Similar to our studies, Wall et al [ 20 ] performed a 6-class classification task but with a random 90–10 train–test split, instead of an 80–20 subject-independent split.…”

“…We subsequently evaluated the Coswara cough (D2), speech (D3), and breathing (D4) datasets. To compare with existing studies [ 20 ], a random 80–20 split was performed for each of these Coswara datasets for model evaluation. Table 16 , Table 17 , Table 18 , Table 19 , Table 20 and Table 21 show the detailed evaluation results and the confusion matrices of the ensemble model for the Coswara cough, speech, and breathing datasets, respectively.…”

“…In comparison with such classification tasks, our ensemble model performed a comparatively more challenging task for the identification of six respiratory abnormalities and achieved competitive performances. Moreover, most of the existing studies, such as [ 13 , 20 , 22 , 23 ], employed a random train–test split, instead of a subject-independent split, which may have audio clips from the same subjects allocated in both training and test sets, although the recordings were collected from different chest locations. In contrast, Zhang et al [ 26 ] utilized a subject-independent split as those used in this research but their work conducted a three-class classification to identify healthy/chronic/non-chronic cases.…”

“…In addition, most studies for the ICBHI dataset, such as Perna and Tagarelli [ 13 ], Wall et al [ 20 ], and Wall et al [ 22 ], adopted RNN models, e.g., LSTM, BiLSTM, or BiGRU, for lung abnormality detection, while a CNN with five convolutional layers and a CRNN model were used by García-Ordás et al [ 24 ] and Zhang et al [ 26 ], respectively. In comparison with these works, our ensemble model incorporates all the above different types of networks (i.e., BiLSTM, BiGRU, CNN, and CRNN), each with PSO-optimized learning hyper-parameters.…”

“…Moreover, Wall et al [ 20 ] utilized BiLSTM and BiGRU with attention mechanisms for lung abnormality classification using the ICBHI and Coswara cough datasets. Sait et al [ 21 ] employed transfer learning based on Inception-v3 combined with MLP for COVID-19 diagnosis using breathing and chest X-ray image inputs, while Wall et al [ 22 ] and Perna [ 23 ] studied BiLSTM and CNN for common lung abnormality diagnosis using breathing audio inputs, respectively.…”