A Narrative Review on Different Novel Machine Learning Techniques for Detecting Pathologies in Infants From Born Baby Cries

This paper reviews the research work on the analysis and classification of pathological infant cries in the last 50 years. The literature review mainly covers the need and role of early clinical diagnosis, pathologies detected from cry samples, challenges in pathological cry signal data acquisition, signal processing techniques, and signal classifiers. The signal processing techniques include preprocessing, feature extraction from domains, such as time, spectral, time-frequency, prosodic, wavelet, etc, and feature selection for selecting dominant features. Literature covers traditional machine learning classifiers, such as Bayesian networks, decision trees, K-nearest neighbor, support vector machine, Gaussian mixture model, etc, and recently added neural network models, such as convolutional neural networks, regression neural networks, probabilistic neural networks, graph neural networks, etc. Significant experimental results of pathological cry identification and classification are listed for comparison. Finally, it suggests future research in the direction of database preparation, feature analysis and extraction, neural network classifiers to provide a non-invasive and robust automatic infant cry analysis model.

How can cry acoustics associate newborns' distress levels with neurophysiological and behavioral signals?

Introduction

Even though infant crying is a common phenomenon in humans' early life, it is still a challenge for researchers to properly understand it as a reflection of complex neurophysiological functions. Our study aims to determine the association between neonatal cry acoustics with neurophysiological signals and behavioral features according to different cry distress levels of newborns.

Methods

Multimodal data from 25 healthy term newborns were collected simultaneously recording infant cry vocalizations, electroencephalography (EEG), near-infrared spectroscopy (NIRS) and videos of facial expressions and body movements. Statistical analysis was conducted on this dataset to identify correlations among variables during three different infant conditions (i.e., resting, cry, and distress). A Deep Learning (DL) algorithm was used to objectively and automatically evaluate the level of cry distress in infants.

Results

We found correlations between most of the features extracted from the signals depending on the infant's arousal state, among them: fundamental frequency (F0), brain activity (delta, theta, and alpha frequency bands), cerebral and body oxygenation, heart rate, facial tension, and body rigidity. Additionally, these associations reinforce that what is occurring at an acoustic level can be characterized by behavioral and neurophysiological patterns. Finally, the DL audio model developed was able to classify the different levels of distress achieving 93% accuracy.

Conclusion

Our findings strengthen the potential of crying as a biomarker evidencing the physical, emotional and health status of the infant becoming a crucial tool for caregivers and clinicians.

Data Collection of Infant Cries for Research and Analysis

Analysis of infants cries may help in identifying the needs of infants such as hunger, pain, sickness, etc and thereby develop a tool or possible mobile application that can help the parents in monitoring the needs of their infant. Analysis of cries of infants who are suffering from neurologic disorders and severe diseases, which can later on result in motor and mental handicap, may prove helpful in early diagnosis of pathologies and protect infants from such disorders. The development of an infant cry corpus is necessary for the analysis of infant cries and for the development of infant cry tools. Infant cry database is not available commercially for research, which limits the scope of research in this area. Because the cry characteristics changes with many factors such as reason for crying, infant's health and weight, age, etc, care is required while designing a corpus for a particular research application of infant cry analysis and classification. In this paper, the ideal characteristics of the corpus are proposed along with factors influencing infant cry characteristics, and experiences during data collection are shared. This study may help other researchers to build an infant cry corpus for their specific problem of study. Justification of the proposed characteristics is also given along with suitable examples.

Neonatal pain analyzer: development and validation

We developed a pain analyzer (ABC analyzer) to perform automatic acoustic analysis of neonatal crying and to provide an objective estimate of neonatal pain. The ABC analyzer uses a validated pain scale (ABC scale) based on three acoustic parameters: pitch frequency, normalized RMS amplitude, and presence of a characteristic frequency- and amplitude-modulated crying feature, defined as "siren cry". Here we assessed the reliability of the analyzer. We enrolled 57 healthy neonates. Each baby was recorded with a video camera during heel prick. Pain intensity was evaluated using a validated scale [Douleur Aigue du Nouveau-Né (DAN) scale] and the analyzer and the two scores were compared. We found a statistically significant concordance between the DAN score and ABC analyzer score (p < 0.0001). The ABC analyser is a novel approach to cry analysis that should now have its properties carefully evaluated in a series of studies, just as is necessary in the development of any other pain measurement tool.

Cry features reflect pain intensity in term newborns: an alarm threshold

The purpose of this study was to assess differences in sound spectra of crying of term newborns in relation to different pain levels. Fifty-seven consecutively born neonates were evaluated during heel-prick performed with different analgesic techniques. Crying was recorded and frequency spectrograms analyzed. A pain score on the DAN (Douleur Aiguë du Nouveau-né) scale was assigned to each baby after the sampling. Three features were considered and correlated with the corresponding DAN scores: 1) whole spectral form; 2) the fundamental frequency of the first cry emitted (F0); and 3) root mean square sound pressure normalized to its maximum. After emission of the first cry, babies with DAN scores >8, but not with DAN scores < or =8 (p < 0.001), showed a pattern ("siren cry") characterized by a sequence of almost identical cries with a period on the order of 1 s. A statistically significant correlation was found between root mean square (r2 = 89%, p < 0.01), F0 (r2 = 32%, p < 0.05), siren cry (r2 = 68.2%, p = 0.02), and DAN score. F0 did not show significant correlation with DAN score in the subset of neonates with DAN scores < or =8 (r2 = 1.4%, p = 0.94), and babies with a DAN score >8 had a significantly higher F0 than those with lower DAN scores (p = 0.016). An alarm threshold exists between high (>8) and low (< or =8) DAN scores: crying has different features in these two groups. When pain exceeds a DAN score of 8, usually a first cry at a high pitch is emitted, followed by the siren cry, with a sound level maintained near its maximum.

A theoretical framework for understanding self-report and observational measures of pain: a communications model

Self-report and observational measures of pain are examined from the perspective of a model of human communication. This model examines the experience of pain as affected by intrapersonal and contextual factors, the process whereby it is encoded into expressive behaviour, and the process of decoding by observers prior to their engaging in action. Self-report measures primarily capture expressive pain behaviour that is under the control of higher mental processes, whereas observational measures capture behaviour that is less subject to voluntary control and more automatic. Automatic expressive behaviours are subject to less purposeful distortion than are behaviours dependent upon higher mental processes. Consequently, observational measures can be used and have clinical utility as indices of pain when self-report is not available, for example, in infants, young children, people with intellectual disabilities or brain damage, and seniors with dementia. These measures are also useful when the credibility of self-report is questioned and even when credible self-report is available. However, automatic behaviours may be more difficult for observers to decode. The model outlined herein takes into account the role of various human developmental stages in pain experience and expression and in understanding the utility of self-report and observational measures. We conclude that both observational and self-report measures are essential in the assessment of pain because of the unique information that each type contributes.