Instrumentation & Measurement Magazine 24-2 - 24

Biomedical Diagnosis of Infant
Cry Signal Based on Analysis of
Cepstrum by Deep Feedforward
Artificial Neural Networks
Salim Lahmiri, Chakib Tadj, and Christian Gargour

he automatic analysis and detection of audio signals is an important field of research with promising
applications in various biomedical engineering problems such as speech, heart murmur, and lung sound analysis
and classification. In this regard, automatic classification of infant vocalizations is becoming an appealing research area for
medical diagnosis in clinical milieu. Indeed, the analysis and
classification of infant cry records is a conventional non-invasive technique to distinguish between healthy and unhealthy
infants.
Recently, various computer-aided diagnosis (CAD) systems have been developed to detect infant pathological cry
records. For instance, an automatic segmentation system for
newborn cry recordings was developed in [1], based on Hidden Markov Models to detect cry expiratory and inspiratory
parts from normal and pathological newborns. The proposed
system achieved 83.79% accuracy. The authors in [2] used a
features vector composed of the prevalence of fundamental
frequency glide, resonance frequencies dysregulation, and
Mel-frequency cestrum coefficients to train a probabilistic neural network. The latter achieved 88.71% accuracy in classifying
healthy and unhealthy records of preterm babies and achieved
67.00% accuracy in classifying healthy and unhealthy records
of full-term babies. More recently, the authors in [3] proposed
an automatic system that combines short-term and long-term
features from different time scales to distinguish between the
cry audio signals of healthy infants from those with respiratory distress syndrome. When trained with Mel-frequency
cepstral coefficients, tilt, and rhythm features, the linear support vector machine yielded to 73.80% accuracy when tested
on expiration samples and 67.80% accuracy when tested on inspiration samples.
In this work, we propose a new CAD system to distinguish between healthy and unhealthy infant cry signals. The
proposed CAD system is composed of four major steps. First,
the original cry signal is pre-processed to remove background

noise and artifacts. This step also includes signal segmentation to differentiate between expiration and segmentation
episodes. Second, the resulting pre-processed cry signal is analyzed to obtain its cepstrum. Third, the obtained cepstrum
coefficients are fed to a deep feedforward neural network
(DFFNN) for training and classification. Fourth, the performance of the cepstrum-DFFNN system is evaluated by
standard classification performance metrics.
The cepstrum is widely employed in audio signal analysis as it provides a description of the spectrum envelope and
spectral richness and characterizes the harmonic and noise
components of the original signal [4]. Moreover, in recent
years, there has been a growing interest in deep learning in
various engineering and science problems thanks to its ability
to extract deep features and achieve high accuracy compared
to existing machining learning techniques. However, the application of deep learning to the problem of infant cry signal
classification for medical diagnosis has not been explored in
the biomedical literature, except in the classification of baby
cry signals under different domestic environment conditions
[5]. In this work, we focus on deep feedforward neural network as it has deeper architectures compared to standard
feedforward neural network, which allow the input data to
be analyzed and transformed multiple times to generate the
output [6]. In this regard, multiple hidden layers make the
DFFNN more appropriate for comprehensive data [7], [8]. In
addition, DFFNN is faster compared to most common deep
learning artificial neural network such as convolutional neural
network and long short-term memory network.
The proposed CAD system for infant cry signal analysis
and classification is shown in Fig. 1, in which the original cry
signal is denoised and segmented. Then, it is processed to obtain its ceptrum signature. The latter is employed to train a
DFFNN which is used to distinguish between healthy and
unhealthy infant cry signals. For comparison purposes, Naïve Bayes, support vector machine, and probabilistic neural

This research is partly supported by the Natural Sciences and Engineering Research Council of Canada (NSERC)
[RGPIN-2016-05067].
24

IEEE Instrumentation & Measurement Magazine
1094-6969/21/$25.00©2021IEEE

April 2021

Instrumentation & Measurement Magazine 24-2

Table of Contents for the Digital Edition of Instrumentation & Measurement Magazine 24-2