Instrumentation & Measurement Magazine 24-2 - 27

we consider an additional experimental protocol where the
data is randomly separated into 50% for learning and 50% for
testing.

Data and Pre-processing
The database is composed of two sets: expiration (EXP) set
and inspiration (INS) set. The EXP set has 2638 cry signals and
INS set has 1860 cry signals. Specifically, there are 1319 healthy
signals and 1319 unhealthy signals in the EXP set. Moreover,
there are 930 healthy signals and 930 unhealthy signals in the
INS set. To record a cry signal, a two-channel sound recorder
with a sampling frequency of 44.1 kHz and a resolution of 16
bits was placed at 10 cm to 30 cm from the infant. The time duration of each recorded signal is within 2 to 3 minutes. Each
original recorded cry signal has been pre-processed to remove
background noise and artifacts. It is also segmented to keep
only respiration and expiration episodes. The segmentation
task is manually performed by using the Wave Surfer tool.
All infant cry signals have been recorded in the neonatology departments of the following hospitals: Sainte-Justine
hospital (Montreal, Canada), Al-Sahel hospital (Beirut, Lebanon), and Al-Raee hospital (Saida, Lebanon). The infants who
entered the study are preterm and full term, and their respective ages range from one to 53 days. The sample includes both
healthy and unhealthy babies and both males and females.
The group of unhealthy babies suffers from various pathologies such as diseases affecting the central nervous system
and respiratory system. Other pathologies include blood disorders, chromosomal abnormalities, and congenital cardiac
anomalies. For illustration purpose, Fig. 3 displays examples
of healthy and unhealthy signals.
Examples of cepstrums representing healthy and unhealthy cry signals are shown in Fig. 4. Recall that the number
of coefficients to be calculated was set to 1000 to allow for a
general representation of the original cry signal, on one hand,
and for fast learning and classification by each single classifier,
on the other hand. According to Fig. 4, INS cepstrums show

Fig. 3. Examples of healthy and unhealthy cry signals.
April 2021

Fig. 4. Examples of cepstrums from healthy and unhealthy infant cry signals.
More variability is observed in cepstrums from healthy and unhealthy signals
associated with inspiration records. In addition, visual inspection of cepstrum
coefficients from inspiration set indicates that those related to healthy infant
cry signals exhibit higher magnitude compared to those associated with
unhealthy cry signals.

more variability compared to those from EXP sets. All experiments are performed by a PC with 1.80 GHz processor and 8
GB installed RAM in Matlab2020b© cloud environment.

Experimental Results
Fig. 5a and Fig. 5b compare the classification results from deep
feedforward (DFFNN), linear support vector machine (SVM),
Naïve Bayes (NB), and probabilistic neural network (PNN)
following 10-fold cross-validation protocol when applied to
EXP and INS sets, respectively. For both EXP and INS sets, the
DFFNN outperforms the linear SVM, NB, and PNN in terms of
accuracy, sensitivity, and specificity.
Specifically, in the problem of classifying EXP cry signals,
the DFFNN, SVM, NB, and PNN achieved an accuracy of
99.92%±0.00, 61.15%±0.04, 58.11%±0.01, and 56.71%±0.01, respectively. In addition, the obtained sensitivity is 99.85%±0.00,
61.03%±0.04, 56.31%±0.01, and 57.70%±0.03, respectively,
for DFFNN, SVM, NB, and PNN. In terms of specificity, the
DFFNN, SVM, NB, and PNN obtained 100%, 61.27%±0.05, and
59.93%±0.02, and 55.72%±0.02, respectively.
Moreover, in the problem of classifying INS cry signals,
the DFFNN achieved perfect accuracy, sensitivity, and specificity. The linear SVM, NB, and PNN, respectively, achieved
an accuracy of 59.57%±0.01, 55.46%±0.02, and 52.63%±0.05, a
sensitivity of 58.82%±0.04, 55.84%±0.01, and 47.10%±0.07, and
a specificity of 60.32%±0.04, 55.08%±0.02, and 58.17%±0.06.
Finally, Table 1 provides the performance measures of each
predictive model when the data set is randomly split into 50%
for training and 50% for testing. As shown, the DFFNN outperforms all classifiers on both EXP set INS sets.
In summary, the DFFNN performs best in distinguishing
between healthy and unhealthy infant cry signals. It is followed by the SVM, and the NB classifier performs the worst. In
terms of processing time for training and classifying EXP and

IEEE Instrumentation & Measurement Magazine 27

Instrumentation & Measurement Magazine 24-2

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