Adaptive Graph Convolutional Neural Network to Detect Asthma Disease

Abstract

This work presents and demonstrates an interpretable graph-based deep learning framework to screen for asthma and four related asthma-respiratory states, directly from lung sound recordings.Introduction/ Importance of Study:Rapid diagnosis is the main component of asthma research because it is part of the avoidable causes of respiratory morbidity that includes more than 262 million people and about 455,000 deaths every year.Novelty statement:Published graph models for respiratory audio fix the adjacency matrix during preprocessing. The proposed Attention-Based Adaptive Graph Convolutional Neural Network (AGCNN) instead treats the adjacency as a constrained learnable parameter refined by gradient descent during end-to-end training, and reads it with four-head attention, so that disease-specific relations between acoustic segments are learned rather than assumed. A systematic comparison against closely related recent studies (Table 1) positions the specific combination of jointly learned topology, multi-head attention and dual-modality interpretability as, to the best of our knowledge, not previously reported for multi-class respiratory sound classification; priority is not claimed over any individual component.Material and Method:The Asthma Detection Dataset v2 used from Kaggle includes five classes (Asthma, Bronchial, COPD, Healthy, Pneumonia) that were band-pass filtered, normalized, and segmented into 25 ms frames; 40 MFCCs per frame constructed the nodes of a Gaussian-kernel k-nearest-neighbour graph (k = 15), and trained three adaptive graph convolution layers under five-fold cross-validation.Result and Discussion:AGCNN achieved 98.18% accuracy (95% Wilson CI 97.26–98.79), 0.9819 weighted F1 score and 0.9993 AUC-ROC, outperforming all the baselines: GraphSAGE (97.05%), GIN (95.71%), GCN (94.30%), GAT (91.99%), SVM (88.89%) and CNN-BiLSTM (85.60%), with an ablation separating the effects of the multi-head attention and adaptive topology, and paired significance testing reported for all baseline comparisons. The recall rates were >0.970 in all classes, including the Bronchial class (which has relatively few subjects), and = 0.993 for COPD. The outperforming accuracy gain was accompanied by clinically coherent reasoning with Grad-CAM finding that the decisions are based on bands of energy within the MFCC that represent wheeze and crackle as well as t-SNE embeddings distinguishing the classes into well separated clusters.Concluding Remarks:In the Conclusions section, Multi-Head Attention enables interpretable, accurate, and CPU deployable respiratory sound classification in the context of learning graphs adaptively from data. All evidence comes from a single openly available database that isnot independent of the dataset, therefore the results do not demonstrate the performance of the algorithm in a clinical diagnostic setting but only algorithmic feasibility for a screening aid, and there is an explicit expectation that the results will be validated with another set of data before it can be used for clinical diagnosis, which is included in the Implications and Recommendations section.