R2REst: A Unified Deep Learning Framework for Estimating Respiration Rate From Respiratory Sounds

Non-invasive respiratory activity assessment, including airflow signal (AF)-derived vital extraction such as respiration rate (RR), tidal volume, expiratory flow rate, etc., and adventitious breathing event detection, are emerging research areas in continuous health monitoring. Recent studies have demonstrated a strong pathological correlation between AFs and respiratory sounds (RSs). In this work, for the first time, we present a unified deep learning framework, namely R2REst, for RR estimation by synthesizing equivalent electrical impedance tomography (EIT)-based AFs from RSs. The proposed framework comprises four major stages: pre-processing, mel spectrogram generation, mel spectrogram-vision transformer-based AF prediction, and lastly, RR estimation by analyzing the frequency spectrum of the predicted AF signal. Unlike prior works that utilizes bio-acoustic modalities other than RSs such as vocal, tracheal sounds in conjunction with pneumotachometry, or capnogram measurements, which are typically cumbersome to collect, our proposed R2REst uses publicly available RSs and AFs from the BRACETS dataset and achieves superior performance, with a mean square error and mean absolute error of 0.001, 0.003, and 0.010, 0.016 (in breaths per minute (BPM)) for tidal breathing followed by deep breathing (TBDB) and cough-speech (TBCS) induced cases, respectively.