Neonatal Pathology Classification from Infant Cry Signals Using a Multi-Branch Graph Neural Network

Even today, the precise and noninvasive identification of neonatal diseases from infant cry signals remains challenging due to the nonstationary and acoustic heterogeneity of cry patterns. This study proposes a graph-based deep learning approach to classify infant cry signals associated with pathologies, specifically Healthy, Sepsis, and Respiratory Distress Syndrome (RDS) categories. Graphs are used to represent cry signals. Nodes are time-acoustic segments with Mel-spectrogram, Mel-Frequency Cepstral Coefficients (MFCC), and Gammatone Frequency Cepstral Coefficients (GFCC) features. Edges are directed temporal links and k-NN links conveying temporal dependencies and non-local acoustic affinities, respectively. A multi-branch approach is introduced to combine heterogeneous feature families with separate projections to a common embedding space before applying graph convolutions. Experimental results show that the proposed Multi-Branch Graph Convolutional Network (GCN) achieves a test accuracy of 95.94% and outperforms the compared models. The macro-average precision, recall, and F1-score are well balanced. Comparisons with previous studies reveal that better performance can be obtained with a convolutional aggregator than with attention-based message passing, which underscores the important role of graph convolutions in capturing infant cry patterns.