Analysis and diagnosis of newborn cry signals based on signal Processing, statistical physics and deep learning

Abstract

Newborn cry is generally due to various conditions related to physiology, pathology, or emotion. In this regard, different patterns in newborn cry signal are associated with health condition. As a result, various computer aided diagnosis systems have been proposed to automatically distinguish between healthy and unhealthy newborn cry signals. Such CAD systems are used to employ specific signal processing techniques combined with machine learning for the analysis and classification of newborn cry signals with acceptable accuracy. The main purpose of our research study is to design new computer aided diagnosis systems based on combination of signal processing and deep learning to improve the accuracy to distinguish between healthy and unhealthy newborn cry signals. In addition, we investigate complexity in such signals based on fractals, entropy, and multifractals to better understand the differences in nonlinear dynamics of cry signals across subjects. For automatic classification of newborn cry signals, we trained various deep learning systems (including deep feedforward, convolution neural networks, and long short-term memory neural networks) with cepstrum-based information. We also used Bayesian optimization method to optimize the hyper-parameters of the support vector machines (SVM) with radial basis function and k-nearest neighbors (kNN), both trained with different audio acoustic features separately or combined which were selected by using a statistical filter. For complexity analysis, we used correlation dimension, approximate entropy, and wavelet leaders. In the task of automatic classification of newborn cry signals based on cepstrum analysis we found that (a) deep feedforward neural network (DFNN) achieved very close to perfect accuracy when applied to expiration infant cry signals and yielded to perfect accuracy when applied to inspiration infant cry signals, (b) DFNN outperformed the linear SVM and the Naïve Bayes systems when tested both on the expiration and inspiration sets, (c) DFNN outperformed very recent works found in the literature, (d) convolution neural networks (CNN) outperformed DFNN and long short-term memory (LSTM) system, and (e) deep learning systems trained with cepstrum descriptors obtained the highest accuracy compared to similar studies in the literature. In the task of classification of healthy versus unhealthy newborn cry signals based on acoustic features we found that (a) the SVM trained with auditory-inspired amplitude modulation (AAM) features achieved the highest accuracy followed by kNN algorithm trained with combination of Mel frequency cepstral coefficients (MFCC), AAM, and prosody, and (b) SVM outperformed most existing works validated on the same database while being considerably fast to perform. In the task of characterization healthy and unhealthy newborn cry signals by using complexity measures, we found that (a) there are significant differences in approximate entropy and correlation dimension across two categories of subjects, (b) healthy infant cry signals show higher approximate entropy level than those of pathological infants, (c) healthy infant cry signals show higher correlation dimension level than those of pathological infants, and (d) healthy signals exhibit a higher degree of multifractality than unhealthy ones. In summary, deep learning systems trained with cepstrum descriptors are promising for analysis and diagnosis of infant cry signals in clinical milieu. Likewise, the nonlinear SVM optimized by using Bayesian optimization and trained by Chi-square based selected features from MFCC, AAM, prosody or combination of those selected features, can be promising for diagnosis of newborns based on their cry signals in clinical milieu. However, deep learning allows achieving the highest performance. Finally, the cepstrum-based approximate entropy and correlation dimension can be considered as biomarkers and could potentially help understanding the physiology of newborn cries and be used for diagnosis purpose.

Date	11 Apr 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Chakib Tadj (Supervisor) & Christian Gargour (Co-supervisor)