Pulmonary function estimation and reliability assessment using hybrid-optimized ensemble learning

Area of Science:

• Respiratory Physiology
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Pulmonary function tests (PFTs) are crucial for assessing ventilatory impairment.
• Limitations of PFTs include accessibility, repeatability, and patient compliance, especially in community and longitudinal settings.
• Investigating non-invasive, low-burden surrogate signals for PFTs is essential for widespread respiratory monitoring.

Purpose of the Study:

• To investigate if acoustic features of voluntary coughs can serve as surrogate signals for spirometry-related indices of ventilatory impairment.
• To develop and validate an ensemble learning framework for modeling spirometry parameters from cough acoustics.

Main Methods:

• A dual-output ensemble learning framework integrating deterministic regression and probabilistic interval estimation was proposed.
• Fifty-eight cough acoustic features and four biometric variables were analyzed using Extreme Gradient Boosting (XGBoost) and Natural Gradient Boosting (NGBoost).
• Hyperparameter tuning was performed using a hybrid Bayesian-Whale Optimization strategy.

Main Results:

• The framework achieved low root mean square errors for forced vital capacity (FVC) and forced expiratory volume in one second (FEV1) (0.20 L and 0.27 L, respectively).
• High coefficients of determination (up to 0.94 for FVC, 0.87 for FEV1) were observed.
• Prediction interval coverage probabilities reached 84.0% for FVC and 77.8% for FEV1, indicating reliable uncertainty quantification.

Conclusions:

• Cough acoustics contain physiologically relevant information for ventilatory impairment assessment.
• The proposed framework enables proxy trend modeling of spirometry indices, not direct replacement.
• Cough-based assessment is feasible for low-cost respiratory screening and longitudinal monitoring, especially when spirometry is impractical.