Mel Frequency Spectrogram Analysis for TB detection on Cough Data
Presentation Type
Poster
Faculty Advisor
Aparna Varde
Access Type
Event
Start Date
26-4-2024 2:15 PM
End Date
27-4-2024 3:14 PM
Description
This study investigates the potential of cough analysis for tuberculosis (TB) detection using a cough audio processing pipeline. Cough data was collected from participants across 7 countries presenting with a cough for at least 2 weeks at outpatient health centers. Demographic information and clinical data were also collected.The proposed methodology involves cough signal extraction, spectrogram generation, and cough sound classification. Initial analysis revealed no significant differences in demographics between TB and non-TB groups when analyzing cough spectrograms. However, key symptoms like weight loss and fever showed promise for TB identification. Additionally, TB-positive subjects exhibited higher heart rates, suggesting its potential as a biomarker. We experimented with VGG16 and ResNet50 deep learning models for image classification using Mel spectrograms, but these models suffered from underfitting and failed to distinguish between TB and non-TB coughs. This suggests that image classification approaches, successful for X-ray image classification, might not be suitable for cough acoustics. Conversely, a 1D convolutional neural network (CNN) trained on Mel-frequency cepstral coefficients (MFCC) signals achieved promising results. For solicited cough data, the model achieved an accuracy of 80.10%, with better precision (83%) for non-TB cough classification. Longitudinal cough data yielded even better results, with an accuracy of 91.01% and balanced precision (91%) and recall (95%) for both TB and non-TB classes. These findings suggest that 1D CNN models show promise for TB detection using cough analysis.
Mel Frequency Spectrogram Analysis for TB detection on Cough Data
This study investigates the potential of cough analysis for tuberculosis (TB) detection using a cough audio processing pipeline. Cough data was collected from participants across 7 countries presenting with a cough for at least 2 weeks at outpatient health centers. Demographic information and clinical data were also collected.The proposed methodology involves cough signal extraction, spectrogram generation, and cough sound classification. Initial analysis revealed no significant differences in demographics between TB and non-TB groups when analyzing cough spectrograms. However, key symptoms like weight loss and fever showed promise for TB identification. Additionally, TB-positive subjects exhibited higher heart rates, suggesting its potential as a biomarker. We experimented with VGG16 and ResNet50 deep learning models for image classification using Mel spectrograms, but these models suffered from underfitting and failed to distinguish between TB and non-TB coughs. This suggests that image classification approaches, successful for X-ray image classification, might not be suitable for cough acoustics. Conversely, a 1D convolutional neural network (CNN) trained on Mel-frequency cepstral coefficients (MFCC) signals achieved promising results. For solicited cough data, the model achieved an accuracy of 80.10%, with better precision (83%) for non-TB cough classification. Longitudinal cough data yielded even better results, with an accuracy of 91.01% and balanced precision (91%) and recall (95%) for both TB and non-TB classes. These findings suggest that 1D CNN models show promise for TB detection using cough analysis.