Predicting Lung Function from Voice Recordings Using Machine Learning: An Optimised Approach

Spirometry is a crucial diagnostic test for pulmonary diseases, where measures including Forced Expiratory Volume in one second (FEV1) have diagnostic and prognostic value. Symptomatic sounds with digital signal processing methods and machine learning techniques have potential for monitoring respiratory disease. Previously, a threshold-based speech and breath extraction method was used to develop a multi-model approach to predict FEV1% predicted from voice recordings. This study further optimises this method to improve predictive performances of models by data augmentation using data segmentation, handling class imbalance, and enhancing predictive models of Random Forest (RF), Logistic Regression (LR) and Support Vector Machine (SVM) with feature selection and hyperparameter tuning. Using 10-second segments with top ten features demonstrates the best results in all three models - regression (RMSE = 9.85), multiclass classification model (Accuracy = 79.28%), and binary classification model with hyperparameter tuning (AUC = 93.57%). The results demonstrate improved performance for the three models after feature selection and hyperparameter tuning. This study presents the limitations of the current experiments and highlights the future work in the ongoing research.