Cough endpoint detection in continuous audio recordings

Cough detection and monitoring plays an important role in the assessment of respiratory conditions in patients. Cough audio can also prove valuable information for models used in the classification and diagnosis of diseases. However, cough detection remains a challenging problem due to variables in audio data often found in environments where cough recordings are captured, as well as the need for disease diagnosis in areas with limited access to internet and devices capable of running deep neural network (DNN) models. This thesis investigates cough endpoint detection using all aspects of machine learning development in an attempt to produce a small DNN model suitable for use in potential mobile applications. The work covers custom dataset design with variations of signal-to-noise ratios, feature extraction to generate both linear and mel frequency spectrograms, low parameter count model design, hyperparameter optimization and finally a detailed evaluation on the achieved performance. Four main classifier architectures are considered including, logistic regression (LR), multilayer perceptron (MLP) and two convolutional neural networks (CNN) with a focus on minimizing model depth and parameter count while maintaining robust cough boundary detection capabilities. Models are trained and evaluated on separate datasets using conventional metrics to assess performance on a classification level, as well as alignment based evaluation methods to assess performance between predicted outputs and ground truth annotations. Hyperparameter optimization identifies optimal parameters among 7290 possible combinations for each classifier. Resulting in optimal models achieving an area under receiver operating characteristic (ROC) curve (AUC) of 0.921, 0.988, 0.985 and 0.992 for LR, MLP, two dimensional CNN and three dimensional CNN respectively. Final testing is done through simulating real world conditions, which presented complications with the original dataset and classifier. After the implementation of post classification processing, final performance achievements showed the classifier detected 99.7% of coughs while 3% of detected coughs were false positives, with false positives primarily caused by audio associated with or similar to coughs sounds. The final results presented a feasible and accurate cough endpoint detection system using compact neural network architecture and provides a strong basis for further research into cough boundary detection and future mobile development.

Grantor dc:publisher

Stellenbosch : Stellenbosch University

Year dc:date.issued

2026

Author dc:creator

• Lamprecht, Christiaan Gerrit

Advisor dc:contributor.advisor

• Niesler, Thomas

Language dc:language.iso

en

Repository record dc:identifier.uri

https://scholar.sun.ac.za/handle/10019.1/136216

OAI identifier oai:identifier

oai:scholar.sun.ac.za:10019.1/136216