Leakage-Free Biomedical Acoustic Signal Classification of Infant Cry for Reliable Neonatal Monitoring

Infant cry analysis is a non-invasive biomedical acoustic signal processing task that can support reliable neonatal monitoring. However, reported high classification accuracy may be affected by dataset imbalance and data leakage rather than true model generalization. This study investigates the effect of leakage-free evaluation and hybrid acoustic feature representation on infant cry classification under imbalanced data conditions. A file-level partitioning protocol was applied before augmentation, feature scaling, and model training to preserve independence between training and testing data. Hybrid acoustic features, including Mel Frequency Cepstral Coefficients (MFCC) statistics, root-mean-square (RMS) energy, zero-crossing rate, pitch-related descriptors, and spectral features, were extracted from eight infant cry categories. Four lightweight machine learning models were evaluated using accuracy, Macro-F1, and balanced accuracy. Under leakage-free evaluation, support vector machines (SVMs) achieved the most balanced performance with 86.41% accuracy, 0.7376 Macro-F1, and 68.43% balanced accuracy. In contrast, leaky evaluation inflated performance above 90% across models. These findings indicate that reliable infant cry classification depends on leakage-free evaluation, imbalance-aware metrics, and biomedical acoustic feature integration rather than model complexity alone.