Investigation of the effects of surgical thyroid disorders on voice characteristics

Postoperative hoarseness is a frequently encountered medicolegal issue in patients undergoing thyroidectomy and represents a significant medical-legal concern. Preoperative acoustic analysis plays a critical role in determining responsibility in such medicolegal cases and in evaluating surgical outcomes, thereby contributing to the improvement of clinical procedures. In the study, 126 patients scheduled for thyroidectomy were first evaluated using videolaryngostroboscopy in the Ear, Nose and Throat clinic. Following this, the /a/ and /i/ vowel sounds recorded from these patients were subjected to acoustic analysis. The results of this analysis were assessed statistically in relation to demographic and clinical features. Subsequently, the acoustic features of healthy and patient voices were used to perform classification with Cubic Support Vector Machines (SVM), Quadratic SVM, Radial Basis Function (RBF) SVM, GentleBoost, and LogitBoost algorithms, and their performances were compared based on accuracy, precision, recall, specificity, Receiver Operating Characteristic (ROC) Area Under the Curve (AUC) and F1-score values. Additionally, SHapley Additive exPlanations (SHAP) analysis was conducted to identify the features contributing most to classification performance. In pre-thyroidectomy cases, a statistically significant increase was observed in features such as jitter, shimmer, and spectral flatness, while a significant decrease was found in Harmonic to Noise Ratio (HNR), spectral flux, peak, power, and energy features (p < 0,05). The effects of nodule size and anatomical localization on acoustic features were demonstrated with statistically significant results (p < 0,05) and high correlation values. All Machine Learning (ML) models showed high performance; however, Cubic SVM was identified as the most successful model in classification, with a 100% AUC value and balanced performance metrics. According to SHAP analyses, the midpower acoustic feature was consistently identified as the most discriminative feature across all models. This study conducted on pre-thyroidectomy cases holds a pioneering position in the literature, as it combines acoustic analysis with ML and statistically demonstrates the relationship between acoustic features and both the anatomical localization and size of thyroid nodules. The findings indicate that voice disorders can be objectively evaluated during the preoperative period. The near-complete differentiation of patient voices from healthy ones may play a critical role in determining medicolegal responsibility in postoperative claims of voice deterioration. Demonstrating the varying effects of nodule localization and size on acoustic features may also be of great importance in monitoring voice disorders in thyroid patients. ML methods, by accurately distinguishing between healthy and pathological voices, show strong potential for integration into clinical decision support systems.