Vowel segmentation impact on machine learning classification for chronic obstructive pulmonary disease.

Area of Science:

• Speech Science
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Non-invasive methods for Chronic Obstructive Pulmonary Disease (COPD) classification are crucial for patient management.
• Vowel-based voice analysis shows promise for assessing phonatory function in COPD.
• Machine learning (ML) models require extensive datasets, necessitating data augmentation techniques like segmentation.

Purpose of the Study:

• To evaluate the impact of segmenting the vowel "a" utterance on the performance of ML classifiers (CatBoost, Random Forest, SVM) for COPD classification.
• To compare the efficacy of full-sequence, segment-wise, and group-wise dataset constructions derived from vowel utterances.

Main Methods:

• Trained CatBoost, Random Forest, and Support Vector Machine models on three dataset types: full-sequence, segment-wise, and group-wise.
• Utilized a nested cross-validation (nCV) approach with grid search for hyperparameter optimization to prevent overfitting.
• Analyzed performance metrics including accuracy, true positive rate (TPR), and true negative rate (TNR).

Main Results:

• The second segment of the vowel "a" utterance, when analyzed segment-wise, outperformed the full-sequence dataset for ML classification.
• The CatBoost model achieved the highest accuracy (97.8% validation, 84.6% test) using the second segment.
• The CatBoost model with the second segment demonstrated the optimal balance of TPR and TNR, indicating clinical utility.

Conclusions:

• Time-sensitive characteristics of vowel production are vital for accurate COPD classification.
• Vowel utterance segmentation can effectively capture these temporal properties, enhancing ML model performance.
• Further research with larger, more diverse datasets is needed to improve generalizability.