Automatic Tuberculosis detection in cough patterns using NLP-style cough embeddings

Coughing is a common symptom of respiratory diseases and past works have used the audio and acoustic properties of coughs to detect those diseases. In this study, we propose a new NLP-style cough embedding technique which is the characterisation of the cough signal as a binary sequence preserving temporal information, and show how this not only allows the discrimination between coughing by tuberculosis (TB) patients and those with other respiratory ailments, but even outperforms more traditional spectral audio features. We have used the vocal audio recordings gathered from 15 TB and 33 non-TB patients, who suffer from other lung diseases. In total, almost 2 hours of cough embeddings were used as feature vectors to train and evaluate four shallow (LR, SVM, KNN, MLP) and two deep architectures (CNN, LSTM) using nested cross-validation. We have also experimentally extracted MFCC, ZCR and kurtosis from the same audio recordings preserving cough patterns and used these to train the classifiers. The results show that an LSTM trained on the cough embeddings achieved the highest AUC of 0.81. When using the audio features, a CNN performed the best by producing the highest AUC of 0.72. We show that detecting TB using cough embeddings preserves privacy and is possible due to the unique temporal cough patterns between TB and non-TB patients. It can also be fused as an additional tool to improve the TB cough audio classification task.