Respiratory Parameter Estimation Using Pharyngeal Phonetics and Machine Learning: Breaking Free from Spirometry

Abstract

The capacity to screen for respiratory diseases is vital for clinical diagnostics. Forced Expiratory Volume in 1 relevant second and Forced Vital Capacity are the two most common parameters or measures of respiratory health. Although respiratory health may be assessed using spirometry or other traditional forms of diagnostics, spirometry has variances in device availability, patient compliance during testing, and complexity of the testing procedure. This study found a novel, non-invasive means of estimating pulmonary function based on the voiced pharyngeal sound of “He” using analysis of the voiced sound. The study explored a model for estimating Forced Expiratory Volume in 1 s and Forced Vital Capacity values based on the outputs from traditional spirometry and features extracted from voice signals. There were a total of 21 features that were extracted from the voiced segments of the pharyngeal sound. All machine learning models of Forced Expiratory Volume in 1 s and Forced Vital Capacity used three machine learning algorithms. Data was collected from 18 male participants, aged 33–49 years old, from June 2022 to August 2022, which resulted in 56 recordings. Among the models that were evaluated in comparison to the linear, neural network, and quadratic models, the quadratic model performed the worst, while the neural network performed the best. The neural network model that used three features estimated Forced Expiratory Volume in 1 s with a mean error of 0.24 % while the two-feature neural network model estimated Forced Vital Capacity with a mean error of 0.58 %. © 2025 Elsevier B.V., All rights reserved.