RESUMEN
Resumen: La evaluación automática de sonidos de auscultación cervical (AC) es una herramienta no invasiva para evaluación de la deglución. Sin embargo, los eventos deglutorios pueden verse enmascarados por fuentes de ruido. Este trabajo propone una metodología de caracterización y clasificación de señales de AC con alta resolución temporal a partir de estetoscopio, para discriminar entre sonidos deglutorios y asociados a ruido. Se adquirieron señales de AC en 10 sujetos sanos durante tres pruebas: toma de líquido, pronunciación del fonema /a/ y aclaramiento de garganta. Se extrajeron características de la señal de AC basadas en coeficientes cepstrales en la escala Mel, transformada wavelet discreta y entropía de Shannon. Las características con mayor relevancia fueron utilizadas como entrada a una máquina de vectores de soporte. Utilizando ventanas de 60 ms - alta resolución temporal - y validación cruzada, se obtuvieron exactitudes del 97.7% para detección de eventos acústicos y 91.7% para sonidos deglutorios. El método propuesto permite clasificación de sonidos deglutorios utilizando estetoscopio -dispositivo común en la práctica clínica- con exactitud comparable a otros trabajos que tienen menor resolución temporal o que utilizan otro tipo de sensores. Este trabajo constituye una primera etapa en el desarrollo de un algoritmo robusto para clasificación de sonidos deglutorios asociados a desórdenes de la deglución, a partir de auscultación cervical, para fines de diagnóstico automático.
Abstract: Automatic evaluation of cervical auscultation sounds (AC) is a non-invasive tool for swallowing assessment. However, the swallowing events could be perturbed by acoustic noise. This paper proposes a methodology of characterization and classification of AC signals acquired by stethoscope with high temporal resolution, in order to discriminate between swallowing sounds and other acoustic noise. AC signals from 10 healthy individuals were acquired with stethoscope during three tasks: liquid ingestion, phoneme /a/ pronunciation and throat clearing. Features based in Mel frequency cepstral coefficients, discrete wavelet transform and Shannon entropy, were extracted. Features with highest Fisher's discriminant ratio were used as input of a support vector machine. By application of 60 ms windows and cross validation, the obtained accuracies were 97.7% for acoustic event detection and 91.7% for swallowing sound detection. The proposed method allows classification swallowing sounds with higher temporal resolution than other works but with comparable accuracy. Furthermore, the use of stethoscope could lead to better acceptation than other sensors by physicians, because it is a common device in clinical practice. This work is a first stage in the development of a robust classification algorithm for sounds in swallowing disorders, oriented to automatic diagnosis.
RESUMEN
OBJECTIVES: The vocal changes after a thyroidectomy are temporary and nonsevere, therefore, obtaining accurate analytical results on the pathological vocal characteristics following such a procedure is difficult. For a more objective acoustic analysis, this study used the cepstral analysis method to examine changes in the patients' voices during the perioperative period regarding sustained vowel phonation. METHODS: The sustained phonation of the five vowels (i.e., /a/, /e/, /i/, /o/, and /u/) by 35 patients with thyroidectomy were recorded by using a Multi-Speech program. Of the 35 patients, 10 were men and 25 were women, with an average age of 51.5 years. Voice data were collected a total of 3 times (preoperatively, 5-7 days after the operation, and 6 weeks after the operation) and were edited according to each fragment (on-set, mid, and off-set) for cepstral analysis. RESULTS: The cepstral analysis on the patients' voices revealed no significant differences between the examination periods of all vowel phonations. However, analysis of the on-set fragment of the vowel /i/ revealed pathological characteristics in which the cepstral measurements of the voice were significantly lower after the operation than before the operation, with the cepstral measurements of the voice increasing further 6 weeks following surgery. CONCLUSION: The results of the acoustic analysis on the on-set fragment of the vowel /i/ will be important data for characterizing the vocal changes during the perioperative period. This study contributes to future research on the mechanisms underlying changes in the voice of patients with a history of thyroid or neck surgery.