Neural network based AI model for lung health assessment

Abstract

Treating pulmonary diseases is pivotal in healthcare since they are the third leading cause of mortality globally. To aid medical experts in diagnosis, various studies have been conducted using artificial intelligence (AI) compatible devices to analyze lung sounds recorded with a stethoscope. In this paper, four datasets have been considered as a combination of two public datasets to assess the performance of the proposed approach. The signals from each dataset undergo a series of pre-processing steps, encompassing normalization, re-sampling, and framing. Thereafter, eight sub-band filters have been taken into account to segregate distinct frequency bands. The sub-band signals are represented using characteristics such as entropy, Lp norm, kurtosis, mean absolute deviation, and standard deviation. This characteristic representation for the signals is then fed to the proposed neural network (NN) for training and classification. The NN architecture consists of three fully connected layers and an output layer for classification. Our proposed approach attains 100% accuracy, specificity, and sensitivity, performing consistently well across all four datasets, which highlights the model’s strong generalizability. The proposed architecture is simple, easy to realize, and has a short training time. The classification outcomes obtained through the proposed NN architecture demonstrate its superiority when compared to the existing methods.