Investigación y desarrollo de técnicas de procesado de señal e inteligencia artificial aplicadas a la recuperación de información biomédica a partir del análisis de señales sonoras respiratorias

Abstract

Obstructive lung diseases are a global health challenge, causing high morbidity, mortality, and economic burden. Auscultation, the primary method for respiratory assessment, is subjective, leading to misdiagnoses and increased healthcare costs. This research focuses on enhancing early detection using a three-phase approach. Firstly, a method combining autoregression-based spectral features and a Support Vector Machine detects crackle events with 80-100% accuracy. Secondly, a cochleogram, modeling human cochlear frequency selectivity, outperforms other time-frequency representations, achieving 85.1% accuracy in wheezes and 73.8% in crackles using Convolutional Neural Networks. Lastly, the Vision Transformer architecture, combined with the cochleogram, demonstrates promise in respiratory sound classification. Despite challenges in standardized databases, evaluations on the ICBHI dataset showcase the effectiveness of the proposed methodology. This research contributes to advancing signal processing and artificial intelligence, aiming to enhance the speed and accuracy of respiratory disease detection, a crucial need in global healthcare.