Artificial intelligence framework for multi-stage lung disease detection with audio signals

Abstract

Automated diagnostic systems are increasingly pivotal in advancing the accuracy and efficiency of medical diagnostics. Due to abnormal changes in human life and pollution, lung disease and cancer cases increasing in huge number. Identification and prediction of lung diseases may help to increase the human life span. This study introduces a robust framework for automatic lung disease detection using respiratory sound signals. The methodology brings together a series of activities like preprocessing, feature extraction, selection, and classification to improve diagnostic accuracy. The adaptive empirical stockwell-transform (AEST) is used to enhance the quality of the signal, whereby extracting and refining features, mainly Mel-frequency cepstral coefficients (MFCC), and Mel-spectrograms, are used. The scalable convolutional geyser network (SCGN) helps to mitigate challenges posed by imbalanced datasets, redundant features, and overfitting, ensuring reliable classification of the features. The model is validated when using the International Conference on Biomedical and Health Informatics (ICBHI) dataset, which validates the performance indicators of the model (F1-score 0.94, accuracy 0.95, precision 0.93, recall 0.94). This is shown superior performance compared to other existing models and demonstrates the framework's ability to diagnose a serviceable and reliable medical diagnosis; which indicates the strengths of combining advances in signal processing and scalable deep learning (DL) in healthcare applications.