Modelos de aprendizado de máquina para a predição do estágio de prontidão para terapia de voz

Abstract

Voice assessment is multidimensional as it addresses various aspects, including the main evaluation methods: perceptual-auditory judgment (PAJ), acoustic assessment, laryngological evaluation, and self-assessment. The latter encompasses the patient’s perspective and provides information on the experience of symptoms, impacts on quality of life, and cognitive-behavioral aspects related to voice. The stage of readiness for voice therapy emerges in this context and reflects the patient’s willingness to adopt changes that contribute to their vocal behavior. Identifying the individual’s stage allows for a more targeted and effective intervention. Moreover, the use of machine learning (ML) models to predict stages of readiness represents an innovation to support clinical decision-making in the evaluation and treatment of dysphonia. The objective of this research is to evaluate and compare the performance of ML models for predicting the stage of readiness for speech-language voice therapy. This is an observational, analytical, and quantitative study with a cross-sectional design, conducted through the retrospective analysis of clinical data, applying predictive ML models to predict the stage of readiness. The data sources for the models were obtained from the initial speech-language and laryngological assessments of patients of both sexes seeking care at the Integrated Voice Studies Laboratory of the Federal University of Paraíba. Data were extracted from the Anamnesis and Vocal Assessment Protocol, the University of Rhode Island Change Assessment – Voice Validated Scale, the Vocal Handicap Index, the Voice-Related Quality of Life Scales, Vocal Symptoms and Vocal Tract Discomfort Scales, and from PAJ, acoustic measures, and laryngological diagnosis. The sample comprised 236 individuals, divided into two readiness stage groups: contemplation and maintenance. Descriptive statistical analysis was performed for quantitative and qualitative variables, and supervised ML algorithms were tested, including logistic regression, k-nearest neighbors (KNN), naive Bayes, decision tree, and random forest. Model performance was evaluated through accuracy, sensitivity, specificity, among other measures. Results indicate that the KNN and Random Forest models performed best with the anamnesis data. For PAJ data, laryngological diagnosis, and dysphonia classification, the Logistic Regression and Naive Bayes models demonstrated balanced performance between sensitivity and specificity. For self-assessment instruments and acoustic measures, the models showed lower performance. In models combining data from different dimensions of voice assessment, KNN stood out, presenting high accuracy. Although still limited in predictive power, these findings suggest that ML models have the potential to contribute to identifying the stage of readiness, enabling more targeted intervention strategies in voice therapy.