Enhancing Lung Sound Classification: A Review of Deep Learning Models with Transform and Spectral Features

Abstract

Accurate identification of  lung  sounds  recorded  by electronic stethoscopes is critical to the early diagnosis of respiratory disorders. In the last thirty years, ML(machine learning) methods have played a significant part in enhancing  the accuracy  of  expert  evaluations.  This  review  article  offers a thorough analysis of the developments in lung sound clas- sification, emphasizing a novel method that makes use of an Enhanced Long  Short-Term  Memory  (ELSTM)  model.  Using a Savitzky-Golay filter to preprocess lung sound audio signals    in order to eliminate noise is the first step. The preprocessed signals are then used to extract important characteristics, such  as Stockwell Transform, spectral-based features, and Short-Time Fourier Transform (STFT). The multi-layered design of the ELSTM model makes use of these characteristics to enhance expressiveness, representation learning, and sequence learning. To improve classification accuracy, a hybrid loss function that combines Categorical Cross-Entropy (CCE) and Focal Loss (FL) is utilized to effectively forecast the error between actual and projected values. This work presents a comprehensive assessment of the literature  on  lung  sound  categorization,  emphasizing  the several approaches and machine learning models that have been investigated. It examines these strategies’ advantages and disadvantages critically while highlighting the ELSTM model’s contributions to solving contemporary problems. This study intends to lead future  research  and  applications  in  the  field  of respiratory sound analysis by integrating prior research and offering a novel model. Ultimately, this will help in the timely identification and treatment of respiratory  disorders.