Crackle and Breathing Phase Detection in Lung Sounds with Deep Bidirectional Gated Recurrent Neural Networks

Benchmarking of eight recurrent neural network variants for breath phase and adventitious sound detection on a self-developed open-access lung sound database—HF_Lung_V1

“…However, the dataset used was relatively small (489 15-s recordings) and the task was to detect inspiratory and expiratory sounds in a 0.1-s segment (time frame) instead of detecting the events of inhalations and exhalations. Messner et al [25] applied the BiGRU to two features, namely Mel-frequency cepstral coefficients (MFCCs) and short-time Fourier transform (STFT)-derived spectrograms, and achieved an F1 score of approximately 86% for breath phase detection based on 4,656 inhalations and 4,720 exhalations and an F1 score of approximately 72% for crackle detection based on 1,339 crackle events. Jácome et al [26] used a faster region-based CNN (Faster R-CNN) framework to obtain a sensitivity of 97.5% and specificity of 85% in inspiratory phase detection and a sensitivity of 95.5% and specificity of 82.5% in expiratory phase detection, which was based on a dataset comprising 3,212 inspiratory phases and 2,842 expiratory phases.…”

“…signals were then processed using STFT [25,54,68]. In the STFT, we set a Hanning window size of 256 and hop length of 64; no additional zero-padding was applied.…”

“…To our best knowledge, almost all previous studies have focused on only distinguishing healthy participants from participants with respiratory disorders and classifying normal breathing sounds and various types of adventitious sounds. Only few studies have reported the performance of sound detection at the recording level using deep learning based on private datasets [24][25][26]. Accurate detection of the start and end times of breath phase and adventitious sounds can be used to derive quantitative indexes, such as duration and occupation rate, which are potential outcome measures for respiratory therapy [27,28].…”

CNN-MoE Based Framework for Classification of Respiratory Anomalies and Lung Disease Detection

“…Feature extraction in state-of-the-art deep learning based systems typically involves generating twodimensional time-frequency spectrograms that are able to capture both fine grained temporal and spectral information as well as present a much wider time context than single frame analysis. While a variety of spectrogram transformations have been utilised, Mel-based methods such as log-Mel spectra [19], [20], [21] and stacked MFCC features [19], [22], [23], [24], [25], [26] are the most popular ones. Some researchers combined different types of spectrogram, e.g.…”

“…A more serious issue with this research field has been the difficulty of comparing between techniques due to the lack of standardised datasets for evaluation. Most publications evaluate on proprietary datasets that are unavailable to others [9], [10], [13], [19], [25].…”