Detecting Childhood Pneumonia Using Handcrafted and Deep Learning Cough Sound Features and Multilayer Perceptron

Area of Science:

• Medical Diagnostics
• Computational Health
• Pediatric Respiratory Diseases

Background:

• Pneumonia is a major cause of child mortality, particularly in resource-limited settings.
• Current diagnostic methods for pneumonia often require specialized facilities.
• Cough sound analysis presents a potential non-invasive method for disease detection.

Purpose of the Study:

• To investigate the efficacy of objective cough sound analysis in differentiating pediatric pneumonia from other acute respiratory diseases.
• To develop and evaluate a machine learning model for pneumonia diagnosis based on cough acoustics.

Main Methods:

• A dataset of 491 cough sounds from 173 children was collected and classified.
• Temporal, spectral, and cepstral features were extracted from cough sounds.
• A multilayer perceptron classifier was trained using extracted features and deep learning embeddings.

Main Results:

• The developed cough sound analysis method achieved 84% sensitivity and 73% specificity.
• The model successfully differentiated between pneumonia and other acute respiratory diseases using cough sounds alone.

Conclusions:

• Objective cough sound evaluation is a viable tool for diagnosing childhood pneumonia.
• This approach can aid in rapid diagnosis, especially in resource-poor regions lacking advanced diagnostic infrastructure.