[Paper Review] Deep Neural Network Based Respiratory Pathology Classification Using Cough Sounds

B T Balamurali, Hwan Ing Hee+6 more|arXiv (Cornell University)|Jun 23, 2021

Respiratory and Cough-Related Research35 references4 citations

TL;DR

This study proposes a deep learning model using bidirectional LSTM (BiLSTM) on Mel Frequency Cepstral Coefficients (MFCCs) to classify cough sounds into healthy or pathological respiratory conditions—asthma, URTI, and LRTI. Trained on a real-world smartphone-collected dataset with clinician-verified labels, the model achieves over 91% accuracy when aggregating multiple cough samples per subject, demonstrating strong potential for mobile-based respiratory screening.

ABSTRACT

Intelligent systems are transforming the world, as well as our healthcare system. We propose a deep learning-based cough sound classification model that can distinguish between children with healthy versus pathological coughs such as asthma, upper respiratory tract infection (URTI), and lower respiratory tract infection (LRTI). In order to train a deep neural network model, we collected a new dataset of cough sounds, labelled with clinician's diagnosis. The chosen model is a bidirectional long-short term memory network (BiLSTM) based on Mel Frequency Cepstral Coefficients (MFCCs) features. The resulting trained model when trained for classifying two classes of coughs -- healthy or pathology (in general or belonging to a specific respiratory pathology), reaches accuracy exceeding 84\% when classifying cough to the label provided by the physicians' diagnosis. In order to classify subject's respiratory pathology condition, results of multiple cough epochs per subject were combined. The resulting prediction accuracy exceeds 91\% for all three respiratory pathologies. However, when the model is trained to classify and discriminate among the four classes of coughs, overall accuracy dropped: one class of pathological coughs are often misclassified as other. However, if one consider the healthy cough classified as healthy and pathological cough classified to have some kind of pathologies, then the overall accuracy of four class model is above 84\%. A longitudinal study of MFCC feature space when comparing pathological and recovered coughs collected from the same subjects revealed the fact that pathological cough irrespective of the underlying conditions occupy the same feature space making it harder to differentiate only using MFCC features.

Motivation & Objective

• To develop an automated, deep learning-based system for classifying cough sounds into healthy or pathological respiratory conditions using real-world audio data.
• To evaluate the performance of a BiLSTM model trained on MFCC features for distinguishing between healthy and pathological coughs in children.
• To investigate the feasibility of using smartphone-collected cough recordings for clinical screening of respiratory pathologies.
• To assess the model's ability to differentiate among multiple specific pathologies (asthma, URTI, LRTI) and identify limitations in multi-class classification.
• To analyze longitudinal changes in MFCC feature space between pathological and recovered coughs to understand feature stability across conditions.

Proposed method

• Collected a new dataset of cough sounds from children in a real hospital setting using smartphones, with clinician-confirmed diagnoses.
• Extracted Mel Frequency Cepstral Coefficients (MFCCs) as acoustic features from raw cough audio signals.
• Trained a bidirectional Long Short-Term Memory (BiLSTM) neural network on MFCC features to classify coughs into two classes: healthy vs. pathological.
• For multi-class classification, the model was trained to distinguish among four classes: healthy, asthma, URTI, and LRTI.
• Applied majority voting across multiple cough epochs per subject to improve prediction reliability and reduce noise from single-sample variability.
• Conducted a longitudinal analysis of MFCC feature space to compare pathological and recovered coughs from the same subjects.

Experimental results

Research questions

• RQ1 Can a deep learning model using MFCC features and BiLSTM architecture accurately classify cough sounds into healthy and pathological respiratory conditions using real-world smartphone recordings?
• RQ2 How does the classification accuracy vary when distinguishing between general pathology (healthy vs. any pathology) versus specific pathologies (asthma, URTI, LRTI)?
• RQ3 What is the impact of combining predictions from multiple cough samples per subject on overall classification performance?
• RQ4 Why does multi-class classification accuracy drop when distinguishing among all four cough categories, and what patterns emerge in misclassification?
• RQ5 To what extent do pathological coughs from different underlying conditions occupy overlapping regions in the MFCC feature space?

Key findings

• The BiLSTM model achieved over 84% accuracy in classifying coughs as healthy or pathological (in general or specific conditions) when trained on individual cough samples.
• When predictions from multiple cough epochs per subject were aggregated using majority voting, classification accuracy exceeded 91% for all three specific pathologies (asthma, URTI, LRTI).
• In the four-class classification task (healthy, asthma, URTI, LRTI), overall accuracy dropped, with pathological classes frequently misclassified into one another.
• However, when considering the broader binary task—healthy vs. any pathology—the overall accuracy remained above 84%, indicating robustness in detecting disease presence.
• Longitudinal analysis revealed that pathological coughs from different conditions (asthma, URTI, LRTI) occupy similar regions in the MFCC feature space, explaining the difficulty in distinguishing them using MFCCs alone.
• The study confirms that smartphone-based cough recording is feasible for clinical screening and supports the development of mobile health applications for early respiratory disease detection.