Soft Wearable Digital Stethoscope Device for Low-Effort Lung Function Estimation via Tracheal Sound and Chest Motion Features Using Machine Learning

Introduction: : Accurate respiratory diagnosis in primary care is often limited by the need for patient effort and clinical expertise inherent to traditional spirometry. This study aims to develop and validate a machine learning-based lung function estimation algorithm using a novel wearable digital stethoscope patch that captures tracheal and chest motion signals during low-effort respiratory maneuvers. By minimizing the reliance on patient performance and clinical supervision, this approach has the potential to enable timely and accurate diagnosis of respiratory conditions in primary care settings.

Materials and

Methods: : The study will enroll 45 healthy adults aged 18–64 with no known respiratory or cardiovascular conditions. Each participant will perform a randomized series of respiratory maneuvers: traditional spirometry, deep breathing (slow inhalation and nasal exhalation), and forced exhalation (sharp exhale without a spirometer), while wearing the patch. Spirometry measurements will serve as the reference. Respiratory phases will be detected in deep breathing and forceful exhalation by aligning local extrema in filtered tracheal sound energy and IMU Z-axis motion signals, with inspiration and expiration labeled based on temporal pairings.

Acoustic and motion features extracted from identified exhalation segments will train and evaluate multiple regression-based machine learning models: k-Nearest Neighbors, Elastic Net Regression, Support Vector Regression, and Simple Linear Regression. The models aim to predict key spirometric parameters such as peak expiratory flow (PEF), forced expiratory volume in one second (FEV₁), and forced vital capacity (FVC). Model performance will be evaluated using metrics such as root mean square error (RMSE), intraclass correlation coefficient (ICC), and Bland-Altman analysis.

Results, Conclusions, and Discussions:: Figure 1A–D demonstrates the complete signal processing pipeline for respiratory phase identification. Figure 1A shows the raw tracheal sound waveform captured during a forced exhalation maneuver. In Figure 1B, the signal is transformed into tracheal sound energy (TSEng) using a short-time windowed variance calculation of a bandpass-filtered (70–1000 Hz) signal. Figure 1C overlays TSEng with the IMU Z-axis signal, which was separately filtered using a 0.01–0.35 Hz Butterworth bandpass filter to reduce motion artifacts. In Figure 1D, respiratory phases: inspiration & expiration, are identified based on the alignment of local minima and maxima from the synchronized TSEng and IMU signals. Subsequent feature extraction is performed on each identified exhalation segment for use in predictive modeling.

Figure 2A–C highlights the transformation of acoustic data into feature representations. Figure 2A shows the raw tracheal signal captured during a forced exhalation maneuver again, while Figure 2B presents its Mel spectrogram, capturing the distribution of energy across perceptually relevant frequency bands over time. Figure 2C shows the corresponding Mel Frequency Cepstral Coefficients (MFCCs), which serve as a compact and informative set of features. Each identified exhalation segment is used to extract mean MFCC values, resulting in a feature matrix for model development.

This study presents a signal processing and feature extraction framework for non-invasive respiratory monitoring using a wearable patch. By integrating tracheal sound energy and IMU motion signals, respiratory phases can be identified without the need for spirometry. Acoustic features, specifically MFCCs extracted from segmented exhalation periods, along with spectral features and biological covariates, will be then used to train supervised regression models including k-NN, Elastic Net, SVR, and Simple Linear regression, to predict spirometric outcomes such as PEF, FEV₁, and FVC. Ongoing work involves collecting additional data and validating model performance on a held-out testing set, using metrics such as intraclass correlation coefficients (ICC), root mean square error (RMSE), and Bland-Altman analysis.

Acknowledgements and/or References (Optional)::