Ensemble deep learning model for dimensionless respiratory airflow estimation using respiratory sound

Abstract

In recent years, computerized methods for analyzing respiratory function have gained increased attention within the scientific community. This study proposes a deep-learning model to estimate the dimensionless respiratory airflow using only respiratory sound without prior calibration. We developed hybrid deep learning models (CNN + LSTM) to extract features from the respiratory sound and model their temporal dependencies. Then, we used an ensemble approach to combine multiple outputs of our models and obtain the respiratory airflow waveform for entire respiratory audio signals as the final output. We conducted a comprehensive set of experiments and evaluated the models using several regression evaluation metrics to assess how the models would perform in various circumstances of different complexity. The methods were developed and evaluated considering respiratory sound and electrical impedance tomography (EIT) data from 50 respiratory patients (15 female and 35 male with an average age of 67.4 ± 8.9 years and body mass index of 27.8 ± 5.6 kg/m2). An external assessment was conducted using an external database, the Respiratory Sound Database (RSD). This was an indirect evaluation because the RSD does not provide the ground truth values of the dimensionless respiratory airflow. In the most complex evaluation task (Task II), we achieved the following results for the estimation of the normalized dimensionless respiratory airflow curve: mean absolute error = 0.134 ± 0.061; root mean squared error = 0.170 ± 0.075; dynamic time warping similarity = 3.282 ± 1.514; Pearson correlation coefficient = 0.770 ± 0.235. External assessment with the RSD showed that the performance of our model decreased when devices different from the ones used for their training were considered. Our study demonstrated that deep learning models could reliably estimate the dimensionless respiratory airflow.