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

Area of Science:

• Pulmonary Medicine
• Biomedical Engineering
• Data Science

Background:

• Spirometry, including Forced Expiratory Volume in one second (FEV1), is vital for diagnosing and monitoring pulmonary diseases.
• Voice and breath sound analysis, using digital signal processing and machine learning, offers a non-invasive approach for respiratory disease monitoring.
• Previous methods predicted FEV1% predicted from voice recordings using a threshold-based extraction method.

Purpose of the Study:

• To optimize predictive models for FEV1% predicted from voice recordings.
• To improve model performance through data augmentation, class imbalance handling, feature selection, and hyperparameter tuning.
• To evaluate enhanced Random Forest (RF), Logistic Regression (LR), and Support Vector Machine (SVM) models.

Main Methods:

• Applied data segmentation and augmentation techniques.
• Implemented strategies to handle class imbalance in the dataset.
• Utilized feature selection and hyperparameter tuning for RF, LR, and SVM models.
• Evaluated models using 10-second voice segments and the top ten features.

Main Results:

• Optimized models demonstrated improved predictive performance.
• Regression model achieved an RMSE of 9.85.
• Multiclass classification model reached an accuracy of 79.28%.
• Binary classification model achieved an AUC of 93.57% with hyperparameter tuning.

Conclusions:

• Feature selection and hyperparameter tuning significantly enhanced the performance of RF, LR, and SVM models.
• The optimized approach shows promise for non-invasive FEV1 prediction from voice.
• Further research is needed to address current limitations and explore future work.