Deep Learning to Classify Respiratory Sounds

2nd place, Digital Health Challenge at the CMG Hackathon

Built a deep learning classifier to differentiate between sick and not-sick patients. Final model - 87.6% accuracy on test set

• Melspectrogram data
• Image Transformations (normalize mean, variance)
• Base: VGG Model pretrained w/ ImageNet
• Top: Densely connected network w/ Dropout
• Optimizer: SGD w/ learning rate decay
• 5 Frozen layers

TODO:

• Ensemble Learning: Combine multiple models prior to densely connected block
• Model Averages: Take models trained on different models, average final results to estimate condition