Asthma Detection Using Respiratory Sound Analysis with Mel-Spectrogram and CNN+Transformer

Asthma is a long-term disease of the airways, and it changes the way a person's breathing sounds.Wheezes and crackles are the clearest examples, which makes recorded breath sounds a sensible starting point for computeraided asthma detection.This paper presents an automated system that labels a respiratory recording as Asthma or Normal.A Normal recording contains no abnormal breath sounds, while an Asthma recording carries features such as wheezes and crackles.Recordings are read in WAV format, and the Librosa library turns each one into a Mel-spectrogram that keeps the frequency content we need.A convolutional network first learns local patterns from this Mel-spectrogram.Its feature maps are then reshaped into a sequence and passed to a Transformer encoder, where multi-head attention relates distant parts of the recording to each other.The encoder output goes through dense layers and a final sigmoid unit, which gives the Asthma or Normal label.For comparison, we also built a plain CNN and measured both models with the same metrics.On the validation set, the hybrid CNN-Transformer model reached an accuracy of 95.18%.We evaluate the models with accuracy, precision, recall, F1-score, sensitivity and specificity, and we also report the confusion matrix, the ROC curve and the AUC to show how well each model separates Normal recordings from Asthma recordings.The system is meant as a research tool.It can support a clinician's judgement, but it cannot replace it.