Voice Analysis and Neural Networks as a Clinical Decision Support System for Patients With Lung Diseases

Abstract

Objective: To analyze the voice of patients with lung diseases, compared with healthy individuals, to detect patterns capable of assessing dyspnea using artificial neural networks (ANNs). Patients and Methods: This research consists of a cross-sectional prospective pilot study performed in a reference tertiary center, which included a group of patients with lung diseases, compared with a group of healthy individuals. Each patient’s voice was recorded in controlled rooms. The following techniques were applied to extract and select signals’ features: statistical analysis, fast Fourier transform, discrete wavelet transform and Mel-Cepstral analysis. In addition, data augmentation was used to avoid overfitting and improve the ANNs accuracy. Results: A total of 195 voices were recorded: 131 from lung disease patients and 64 from healthy individuals, separated according to gender and age. Using data augmentation, 751 additional audio samples were generated: 501 from healthy individuals and 445 from patients with lung disease. Among male participants, 133 samples were related to lung diseases and 197 were related to healthy ones. From them, 264 audios were used for ANNs training, obtaining an accuracy of 89%. In the female group, 312 had lung diseases and 304 were healthy. Among them, 492 audios were used for training, resulting in an accuracy of 87.6%. Conclusion: Spectral analysis techniques applied to voice recordings using ANNs have reported high accuracy in the efficient diagnosis of lung diseases when compared with healthy individuals.