Voice dysfunction is a common complication following thyroid surgery. However, the application of explainable machine learning for predicting postoperative voice recovery remains largely unexplored. Therefore, an investigation was done to examine voice recovery based on acoustic, objective, and glottal features. Voice recordings were collected from female patients before surgery and one month after surgery. Acoustic and glottal parameters, including Quasi Open Quotient, Speed Quotient, age, and others, were automatically extracted from the recordings. Random Forest, Support Vector Machines, and Logistic Regression with Sequential Feature Selection were applied to examine model behavior and identify feature importance. Model stability and interpretability were evaluated across cross-validation folds. Performance metrics varied over folds, highlighting the exploratory and statistically fragile nature of predictions in small datasets. SHAP (SHapley Additive exPlanations) analysis revealed variability in feature contributions, emphasizing the need for cautious interpretation and detailed methodological reporting. Our findings provide preliminary guidance for applying explainable machine learning to small biomedical datasets. They demonstrate the importance of careful methodological design.
Haddou et al. (Thu,) studied this question.