Time series optimization neural network model for identifying abnormal lung sounds in children

Abnormal lung sound auscultation is an important tool for diagnosing bronchopulmonary diseases in children. Addressing the issues of high computational resource demands and low recognition rates commonly associated with spectrogram image recognition methods used in the classification of children's abnormal lung sounds, this paper proposes a hybrid model combining Mel-scale Frequency Cepstral Coefficients (MFCC) features, Convolutional Neural Network (CNN), and Bidirectional Long Short-Term Memory (BiLSTM) network for classifying abnormal lung sounds in children. This method uses CNN to extract spatial features from MFCC, and BiLSTM to capture the temporal characteristics of the MFCC audio features, thereby establishing the BCNnet (BiLSTM-CNN network) model. This paper collects and establishes a dataset of children's lung sounds. On this dataset, the proposed method achieves an average accuracy of 75.3%, representing a 3.7% improvement in accuracy compared to the CNN (parallel-pooling) model that uses spectrograms as input. Additionally, the proposed model demonstrates improvements in both size and recognition speed.