Selección de un método de aprendizaje automático para clasificar patrones biomarcadores de lesiones precancerosas de las cuerdas vocales

Abstract

The National Survey on Drug, Alcohol and Tobacco, 2016-2017, notes that 15.6 million Mexicans are active smokers and, by 2030, expect the death of 8 million cancers of the larynx or lung. Therefore, the World Health Organization (WHO) recommends detecting precancerous lesions of the larynx. This is possible, as they are characterized by a biomarker pattern manifested by the alteration of the biomechanical interpretation of the vocal cords, regardless of the sex and age of the smoker. The goal of this article is to evaluate three machine learning methods: neural networks, Gaussian networks, and decision tree to determine the method that best solves the problem of detecting patterns of precancerous vocal cord injury biomarkers. It uses the WEKA tool and a knowledge bank, endorsed by NOM-012-SSA3-2012, with 250 patterns, and provided by the Luis Guillermo Ibarra National Institute of Rehabilitation, Ibarra. The performance of the methods was compared by ROC curves and confusion matrices, under the criteria established by ISO-5725. The decision tree the method that best responds to the detection of patterns of biomarkers of precancerous lesions of the vocal cords.