Automated Detection and Classification of Newborn Crying with Machine Learning

This study aims to predict the crying behaviors and needs of newborn babies using machine learning (ML). This study also aims to support perceptual analysis with a fully ML method to evaluate the shape of newborn crying in a time band. Real-time data were obtained from 32 term newborns in stable medical condition in the neonatal intensive care unit, collected over 12-hour periods, and ML was used to estimate the needs associated with their crying behavior. In each recording, crying units are detected, even if they are long-lasting, and their classification is performed according to six basic behavioral patterns (resting, hungry, sleepy, pain, burp, and distress). While determining these behaviors, noisy signals other than the target signals are filtered from the system, contributing to the acceleration of the analysis process. This method is tested on synthesized signals and applied to recordings from selected newborns. The audio signals received from the newborn baby are transformed into a spectrogram image using the short-time Fourier transform technique. In the latter, the deep convolutional neural network technique takes the spectrogram images as input, and the baby's behavior is obtained from the convolutional neural network and passed to the support vector machine classifier. The results obtained within the scope of this study were compared independently of the recorded voice analysis and the environment that meets the needs of the newborn baby, and approximately 91,856% similarity was obtained through cry unit band time. The system designed in this study is contactless and inexpensive, making it quite suitable for routine clinical practice and may contribute to health professionals' ability to meet their babies' needs.