Infant Health Risk Detection Using Deep Learning-Based Cry Analysis for Classifying Normal Infant Cry Signal

Abstract: Infant cry signals carry critical diagnostic information regarding a newborn’s physical state, discomfort, and health risks. While parents rely on intuition, manual identification of specific cry reasons is highly subjective and error-prone. The paper proposes an automated framework for infant health risk detection using deep learning-based cry signal analysis. This paper focus on the classification of normal infant’s cry signal into 5 different classes: Neh for Hungry, Eh for burp or Pain, Owh for Sleepy, Eairh for Internal pain or Tummy pain and Heh for discomfort. By applying DSP (Digital Signal Processing), we extract a bonded acoustic feature containing MFCCs (Mel-Frequency Cepstral Coefficients) and pitch (F0) parameters. A convolutional neural network (CNN) output is match with traditional K-Nearest Neighbors (KNN) and support vector machine (SVM) models. The paper proposes a CNN model, which attains good accuracy in multi-class cry classification. The trial evaluation for the classification is conducted on the Kaggle Infant Cry Audio Corpus and the ICSD dataset using Python with TensorFlow, PyTorch. The experimental findings indicate that the combination of MFCC features with pitch improves the classification accuracy and offering a reliable tool for the caregivers and the pediatricians.