AHP-CM: Attentional Homogeneous-Padded Composite Model for Respiratory Anomalies Prediction

Early diagnosis, treatment and regular monitoring can limit the spread and adverse effects of respiratory diseases. Shortage of trained physicians is one of the main obstacles to ensure early diagnosis and treatment which can be overcome by making lung auscultations automated. To automate lung auscultations and to identify anomalies like crackles, wheezes and/or both, in this work, we propose a hybrid deep learning model combining a Convolutional Neural Network (CNN) model, ResNet34 as a feature extractor, and a Long Short-Term Memory (LSTM) as a predictor, along with a novel augmentation technique called homogeneous padding over the ICBHI-2017 dataset. We have also added an attention layer in the feature extractor to allow the model learn the important region of the feature vector. The proposed model has outperformed the recent state-of-the-art models in this regard. We have also found that the inclusion of the attention layer, and the LSTM as a predictor has improved the performance of our model in 2-class and 4-class anomaly predictions.