Investigation of voice changes before thyroid surgery with machine learning methods

Abstract

Objective: The study aims to understand the clinical significance of these changes and the potential impact of surgical interventions by examining common vocal changes before thyroid surgery. Material and Methods: The study had a prospective design and data sets were created from the Saarbruecken voice dataset and the “a” and “i” sounds of 126 individuals (34 males, 92 females) in two groups of healthy and thyroid surgery candidates obtained from our clinic, and then processed with signal processing methods and extracted features. These features were used to classify healthy and sick subjects using support vector machines (cubic and quadratic), k-nearest neighbor (k= 5 and k= 7) and ensemble learning (gentleboost and bag) classifiers. After the attributes that were effective in the classification of healthy and patient groups were determined by Shapley value method, Kruskal-Wallis H test with Post Hoc Tamhane’s T2 test, MannWhitney U Test and Spearman correlation test were used to evaluate the changes in attributes. Results: The laryngoscopies of the thyroid surgery candidates included in the study performed in the ENT clinic were normal. All machine learning methods achieved 94.40%-100.00% success in healthy-diseased group discrimination. Shapley analysis identified mid-frequency power, spectral entropy, formant-1 energy and formant-2 energy as effective features for healthy-diseased discrimination; mid-frequency power, formant-2 bandwidth and formant-2 for diseased-healthy “a” sounds discrimination; formant-1, formant-2 bandwidth, formant-2 and mid-frequency power as effective features for “i” sound discrimination in diseased and healthy groups. According to this evaluation, these attributes are effective in the discrimination of healthy-diseased, “a” sounds and “i” sounds. The selection of these attributes is important for improving classification performance and obtaining accurate results. In addition, statistically significant associations (p< 0.05) were found between effective attributes and presence of nodules, sex, smoking, and thyroid gland volume. Conclusion: All patients who are candidates for thyroid surgery have a voice disorder, and it is recommended that patients should be evaluated for voice disorder during their previous follow-up and treatment. It is observed that machine learning methods are effective in these evaluations.