Use of prosodic features in infant cry diagnostic system

The newborn’s Cry Audio Signal (CAS) is made up of a rhythmic sound. Imagine that the newborns would not cry; in this case, we had no way of understanding the newborn’s needs. Needs like hunger, pain, illness, or just the need to hug. When a parent hears the sound of a newborn crying, stress hormones are released into the parent’s body, which leads to high blood pressure, heart rate, and muscle tension, and thus the parent tries to stop crying by alleviating the newborn. Crying is explained as a graded signal that is a stimulus in the behavioural system. Newborns can elicit the surrounding people’s reaction by crying, so newborns’ crying is regarded as an early behaviour for survival in the behavioural system. \n \nThe cry-researchers found the newborns’ CASs having concealed information about the newborn’s physical and psychological states. The newborns’ brain changes the amount of traction in the vocal cords through the cranial nerves. Because the cranial nerves control crying, the cry-researchers made a connection between crying and the brain. The research on newborns’ CAS to investigate the potential of discriminating characteristics started in the 1960s. It started with the subjective auditory investigations, and interestingly, several reports showed that mothers and the hospital staff often could distinguish the needs of newborns only by listening to them. The investigation was then followed by time, frequency, and spectrographic domains analyses. Through these examinations, distinctive patterns were revealed that determine group characteristics. Finally, to avoid the tedious task of analyzing a large amount of information in newborns’ CASs by humans, automated machine-based analysis was proposed. Such a system for analyzing newborns’ CASs can considerably speed up the investigation time and automatically classify them. This is where machine learning models were introduced to capture the statistics in the newborns’ CASs. \n \nThis thesis aims to develop the Newborn Cry Diagnostic system (NCDS) to automatically identify sick infants’ CASs from healthy ones without any newborn physical examination. An NCDS includes three main stages of preprocessing, feature extraction, and model training for classification. This research presented here explores patterns at different levels of newborns’ CASs in the feature extraction phase. The analysis includes investigating the short-term and long-term information in the newborn’s CASs for potential pathologically informed features. Our main contribution in this work is the use of the prosodic features to investigate the long-term statistical patterns in newborns’ CASs. We explored the effectiveness of rhythm, tilt, and intensity feature sets in NCDS. The prosodic feature sets of tilt and rhythm have never been studied in NCDS. The high-level information, namely prosodic features, was found to improve the discriminative ability within audio signals in speech and language recognition systems. \n \nRegarding the short-term feature sets, the common feature set successfully examined in NCDS is Mel Frequency Cepstral Coefficients (MFCC). Another innovation of this work is that we employed the short-term feature set of Auditory-inspired Amplitude Modulation (AAM) for the first time in the NCDS. Our goal was to compare the functionality of the AAM feature set in NCDS with the most influential examined feature set of MFCC and explore the fusion potential of this feature set with MFCC and the prosodic feature set. \n \nThe performance of each feature set was evaluated using a collection of classifiers, including support vector machine, decision tree, perceptron neural network and discriminant analysis. We also examined the majority voting method to upgrade the classification results, which has not previously been reported in the literature relating to developing an NCDS. \n \nOur study primarily focused on two critical pathologies of respiratory distress and sepsis, ranking as the 11th and sixth leading causes of death in Canada. In the end, we came up with a comprehensive model encompassing 34 pathologies common among newborns.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.