Convolutional Dense Neural Network Based Spirometry Variable FVC Prediction Using Sustained Phonations

Abstract

Spirometry is a lung function test used to diagnose and monitor lung diseases like asthma, pneumonia, chronic obstructive pulmonary disease, etc. Spirometry measures forced vital capacity (FVC), forced expiratory volume in 1 sec (FEV1), and their ratio to determine lung health. Spirometry is very time-consuming, strenuous, and requires proper training. Alternate methods based on voice for diagnosis and monitoring of lung health are promising because they are faster, easy to do, and require minimal training. Non-speech sounds, namely, cough and wheeze, have been used to predict spirometry variables, but the role of speech sounds that occur in natural speaking for a similar task has not been explored. In this work, the spirometry variable, FVC has been predicted from sustained phonations of vowel sounds using a convolutional dense neural network (CDNN). Mel-spectrogram has been used as a feature. An experiment conducted using 160 subjects indicates, /i/ is the best sound and /u:/ is worst for the prediction task with an average Mean Absolute Error of 0.67l(±. 07l) and 0.70l(± 0.13l) among all sustained phonations of vowels sounds considered in this work.