What question did this study set out to answer?

The aim is to enhance the classification of lung sounds to identify respiratory disorders earlier and reduce subjectivity in auscultation.

February 8, 2026Open Access

Improving Lung Sound Classification through Multi-Transform Features and Enhanced LSTM Modeling

Key Points

The aim is to enhance the classification of lung sounds to identify respiratory disorders earlier and reduce subjectivity in auscultation.
Developed a classification framework using a mixture of multi-transform spectral features.
Implemented Enhanced Long Short-Term Memory (ELSTM) network with a specific architecture.
Used the ICBHI 2017 database containing records from healthy and pathological subjects.
Applied Savitzky Golay filter to reduce noise in lung sound recordings.
Combined Categorical Cross-Entropy and Focal Loss during training to address class imbalance.
Achieved a sensitivity of 93.7% and a specificity of 94.2%.
Outperformed baseline models like CNN, GRU, and Bi-LSTM in terms of sensitivity and accuracy.

Abstract

Timely and correct perception of the lung sounds is necessitated in order to identify any respiratory disorders early enough as well as trimming down the subjective nature in the exercise of the conventional auscultation. In the current paper, a classification structure based on a hybrid loss incorporates a mixture of multi-transform spectral features and an Enhanced Long Short- Memory with Train (ELSTM) network. The test was run on the ICBHI 2017 database of 920 records of 126 subjects including healthy and pathological cases. A Savitzky Golay filter has been used to eliminate the noises to improve the quality of the signal. The Short-Time Fourier transform (STFT) features, Stockwell transform features, spectral roll-off features were combined and fed into a three layer ELSTM dropout network (128-64-32 units). A combination of the Categorical Cross-Entropy and Focal Loss was adopted in training so as to handle the issue of the class imbalance. The sensitivity of the system is 93.7 and the specificity of the system is 94.2 which is higher than CNN, GRU and Bi-LSTM baselines which attained 96 percent accuracy and 93.7 percent sensitivity.

Bookmark

View Full Paper