Respiratory rate estimation from breath sounds based on deep learning

Respiratory rate (RR) is a fundamental vital sign for evaluating pulmonary and cardiovascular function, with demonstrated value in predicting short-term mortality and early clinical deterioration. As one of the most sensitive physiological indicators, RR often reflects physical and psychological imbalances prior to alterations in other vital signs. This study aims to develop a robust deep learning (DL)-based framework for RR estimation from respiratory sounds. A dedicated database for RR estimation was created, encompassing a wide range of simulated and real-world acoustic environments. Two data augmentation techniques, commonly used in audio processing but novel for this biomedical task, were applied to expand the size of RR breathing sound signals. The framework leverages pre-trained convolutional neural network (CNN) architectures for feature extraction from spectrograms, which are then used to estimate RR. The proposed method was evaluated against several state-of-the-art approaches, encompassing envelope-based methods, respiratory spectral activation models and neural network architectures such as CNNs and recurrent neural networks (RNNs). In noisy acoustic conditions, ResNet50 outperformed all other approaches, achieving a mean absolute error (MAE) of 0.63 breaths per minute (bpm) even at a signal-to-noise ratio (SNR) of –20 dB. It was followed by the LSTM-based method (2.35 bpm) and the CNN-based method (3.12 bpm). Our findings reveal the limitations of existing approaches under severe acoustic degradation and underscore the robustness of ResNet50, which maintains low errors across RR ranges, noise levels, and all evaluated acoustic environments.