Content uploaded by Edward Sazonov
Author content
All content in this area was uploaded by Edward Sazonov on Oct 12, 2015
Content may be subject to copyright.
Ob je ct iv e M on it or ing o f In fant Fe ed in g Beh av io r us ing a J aw Mo ti on S ens or , M uh am m ad F ar ooq , Pa ula C C ha nd le r -L an ey ,
Ma ri a He rna nd ez -Re if a nd E dwa rd S az ono v, J o ur na l Of H e al th ca re E ngi ne er in g, Vo l. 6, N o. 1 , p p. 23-40, Marc h 2015.
ht tp :/ /dx.d oi .o rg /1 0.1 26 0/ 20 40 - 229 5. 6. 1. 23. Cor re sp on di ng au th or : E. Sa zo no v, es az ono v@ en g. ua.ed u, + 1 (2 05) 348 -1981
Objective Monitoring of Infant Feeding Behavior
using a Jaw Motion Sensor
Muhammad Farooq1, Paula C Chandler-Laney2, Maria Hernandez-Reif3 and
Edward Sazonov1
1Department of Electrical and Computer Engineering, University of Alabama,
Tuscaloosa, AL USA
2Department of Nutrition Sciences, University of Alabama at Birmingham,
Birmingham AL USA
3Departmnet of Human Development & Family Studies, University of Alabama,
Tuscaloosa, AL, USA
ABSTRACT
Rapid weight gain during infancy increases the risk of obesity. Given that infant feeding may contribute
to rapid weight gain, it would be useful to develop objective tools with which to monitor infant feeding
behavior. This paper presents an objective method for examining infant sucking count and sucking rate
during meals. A piezoelectric jaw motion sensor and a video camera were used to monitor jaw motions of
10 infants during a meal. Videotapes and sensor signals were annotated by two independent human raters,
marking individual sucking bouts and number of sucks in 10s epochs. Annotated data were used as a gold
standard for development of the computer algorithms. The sensor signal was de-noised and normalized
prior to computing per epoch suck counts/rate. A leave-one-out cross-validation scheme resulted in a
mean absolute error rate of 18.52% and an average intra-class correlation coefficient value of 0.86
between the human raters and algorithm.
Keywords: sensors, obesity, sucking rate, feeding behavior, jaw sensor, breast-fed, bottle-fed, infants.
1. INTRODUCTION
Obesity is caused by the excessive accumulation of body fat. Previous research suggests that infancy
may be a critical period during which life-time risk for obesity develops. Indeed, bottle feeding [1,2] and
rapid weight gain during the first six months of life are associated with subsequent life-time risk for
obesity [3,4,5,6,7,8]. In addition, bottle-fed infants, versus those fed from the breast, have different
energy intake, meal patterns, and within-meal behavior [9,10]. Furthermore, irrespective of the feeding
mode, vigorous sucking at 3 months of age is positively associated with adiposity at 12 months of age.
Despite the fact that excess energy intake during infancy is a primary determinant of rapid weight gain
and subsequent body size [14]; methods to monitor feeding behavior and to assess infant energy intake
are limited. Consequently, the study of infant feeding behavior may yield valuable insight into some of
the earliest feeding patterns that contribute to later obesity.
Monitoring of sucking can provide a better understanding of the feeding behavior in both breast-fed
and bottle-fed infants. Previous research has suggested the use of sensor modalities such as ultrasound
[11,12] and intraoral pressure measurement [12] for monitoring of sucking, swallowing and breathing
movements. These methods provide a close estimate of sucking but their practicality in free living
situations for extended periods of time is limited. Within-meal feeding behaviors, such as sucking bursts
and pauses in sucking, have previously been assessed by videotape and/or with an apparatus containing a
pressure transducer attached to the breast or bottle nipple [10,13,14], but neither procedure is practical for
the assessment of multiple meals over the course of one or more days. Consequently, the need exists for
the development of simple and objective techniques to estimate number and rate of sucking and monitor
feeding behavior for both breast-fed and bottle-fed infants.
This paper presents a novel approach for monitoring infant sucking count and sucking rate during
meals. A similar procedure using a jaw motion sensor has previously been found to provide accurate and
objective monitoring of ingestive behavior in adults [15,16,17,18,19]. Since infants use their jaws during
sucking (both bottle-fed and breast-fed), there is potential to use the jaw motion sensor to capture sucking
count and rate during feeding episodes. According to the authors’ knowledge, no one has previously
reported the use of a jaw sensor to monitor feeding behavior in infants.
The goal of this study was to evaluate the technical feasibility of using a non-invasive jaw motion
sensor for monitoring sucking count and thereby sucking rate of infants. A comparison of human raters
and results obtained using jaw motion sensor is presented to show that the jaw motion sensor with related
signal processing algorithm is able to provide an accurate estimation of sucking rate and sucking count for
both breast-fed and bottle-fed infants.
2. METHODS
2.1 Participants
Ten infants were recruited into this study. Infants were eligible for inclusion if they were at least 2.0
months of age and less than 5.0 months of age and were healthy, with a current weight-for-length not less
than the 5th percentile based on Centers for Disease Control and Prevention (CDC) growth curves [20].
Infants were excluded if they were younger than 37 weeks gestation at birth and/or weighed less than
2500g at birth. The Institutional Review Boards (IRBs) at The University of Alabama Tuscaloosa (UA)
and The University of Alabama Birmingham (UAB) approved these studies. Mothers of infants provided
informed consent before the start of the experiment. Table 1 shows the details of all subjects.
Table 1 Information regarding the participant infants
Attribute
Value
Race (Caucasian/African American/Other)
6/1/3
Number of female infants
5
Number of male infants
5
Number of Breast fed infants
6
Number of Bottle-fed infants
4
Gestational age at delivery (weeks)
39.9 ± 1.5
Birth weight (kg)
3.6 ± 0.3
Average age (weeks)
17.91 ± 2.54 weeks
Range of age (weeks)
13.9-21.1 weeks
2.2 Protocol
Participants and their mothers came to the UAB Child Health Research Center (CHRU) located within
Children’s of Alabama, to complete the study protocol. Each infant came to the laboratory for one visit,
and visits were timed to coincide with when the infant was expected to be hungry, and not less than 2
hours following a previous meal. After informed consent was obtained, infant weight (without clothing)
and length (supine) was measured using standard clinical procedure. Infants then underwent a weighed,
timed, and videotaped meal test while wearing the jaw motion sensor, following which infant body
composition was assessed by air displacement plethysmography (PeaPod®; Cosmed Inc., Concord, CA).
2.3 Meal test
All infants were weighed to the nearest 0.1 gram while wearing a clean diaper only, on the weighing
scale of the PeaPod ®. For bottle-fed infants, the prepared bottle was also weighed to the nearest 0.1 gram
prior to the meal test. The mother sat in an arm chair to breast or bottle feed her infant. The mother held
her infant in a side-lying position, with the infant’s head supported in the crook of her bent arm. The jaw
sensor (described below) was adhered directly under the ear of the infant, behind the jaw, on the side of
the face that would face away from the mother during the meal. The longitudinal axis of the sensor was
perpendicular to the longitudinal axis of the ear. Two video cameras (Samsung HMX-F80) were
positioned to record mouth and jaw movements immediately prior to and during the feed; one focused
from the side and the other from above the infant’s head. Mothers were instructed to begin the meal once
the video and jaw sensor were recording, and to make the meal as “natural as possible”. Mothers could
interrupt the meal to burp the infant or to change positions, as needed, and recording continued during this
time. Meals ended when the infant fell asleep, finished the bottle (if applicable), refused to consume
more, and/or mothers’ indicated that the infant was finished feeding. After burping, infants were
reweighed on the PeaPod® scale wearing the same diaper as they were weighed in prior to the meal and
bottles were also reweighed as appropriate. Any milk lost through spillage or regurgitation was captured
in pre-weighed burp cloths so that milk consumption could be corrected for spillage.
2.4 Body composition assessment
Body composition was measured by air displacement plethysmography (PeaPod®; Life Measurement
Instruments, Concord, CA). In brief, after infant length was measured, infants were weighed on the
PeaPod® scale while wearing only a tightly-fitted stocking cap to compress hair, and then placed inside
the test chamber where infant volume was assessed. Infant fat mass and fat free mass was then calculated
using a two compartment model [21].
2.5 Training for coding discrete infants feeding behaviors: Sucking count
A researcher with expertise in coding infant behaviors trained two human raters to code the videotapes
of the infant feeding sessions. Each human rater met individually with the researcher for the training. The
human raters were informed that they were assisting with testing the reliability of a jaw motion sensor
designed to record feeding behaviors. At the training session, video segments of infants being breast-fed
and bottle-fed were displayed while the researcher pointed out discrete examples of the two behaviors the
human raters would be asked to code: infant sucks and sucking bursts. Each behavior was discussed and
reviewed separately. Only the data for infant sucks (discrete suck) are discussed in this paper.
A discrete suck was operationally defined as one down and one up jaw movement with the infant’s lips
wrapped around the mother’s or bottle’s nipple. In other words, one discrete suck consisted of two
movements: one down followed by one up jaw movement. The researcher counted, out loud, numerous
discrete sucks while watching segments of the videotapes with the rater. Afterwards, the researcher asked
the rater to count discrete sucks out loud for segments of breast and bottle fed infant videos. After several
trials counting discrete sucks out loud, the researcher and the rater viewed several new video segments of
an infant being breastfed, and then separately of an infant being bottle-fed. During the viewing of each
new segment, the researcher and rater independently and quietly counted the frequency of sucks and
recorded the tallies on paper. At the end of the coding, the tallies were examined. A criterion of 90%
agreement with the researcher was established for the rater to be deemed reliable in counting discrete
sucks.
2.6 Coding of videotapes
After the training, video recordings and time-synchronous jaw sensor signals were annotated by the
two trained raters using a modified version of the software introduced in [19]. To provide synchronization
between the sensor signal (described below) and the video, data collection was started on the Android
phone when it was in the view of one of the cameras. To maintain homogeneity of the scoring process, a
protocol similar to other studies [22, 23], was developed to mark meal initiation and termination, and
sucking count to code the videos. Human raters marked the start and end of the meal in the video. The
period in the video between the start and end of the meal was divided into segments of 10s called epochs
(a total of epochs), in the scoring software and sucking count was computed for each epoch. The time
sampling unit of 10s intervals was agreed upon because in the pilot coding it was deemed to be a
manageable period to count discrete sucks. The time sampling units also provided cluster data for
conducting intra-class correlation (ICC) analyses between the human raters, and also the human raters and
the jaw sensor algorithm. Epochs which contained partial intake (video segments where jaw was not
visible in the video) or no intake were discarded. For epoch, the average values of both human
raters gave the annotated per epoch sucking count denoted by and annotated per epoch sucking
rate was computed as . There values were used as gold standard for algorithm
development.
2.7 Jaw motion sensor
The sensor data collection system consisted of a jaw motion sensor, a data acquisition device and an
Android smart phone. The sensor system was originally designed to be used in adults [16] and used in this
experiment without modifications. The jaw motion sensor was a piezoelectric film element (DT2-028K)
(Measurement Specialties Inc. VA). Jaw movements during the sucking process bend the sensor and
create an electrical signal proportional to the amount of bending. Signal from the sensor was buffered and
amplified using an operational amplifier circuit and then digitized by a microcontroller of an Automatic
Ingestion Monitor (AIM, [16]) at a sampling frequency of 1 kHz. Digitized signals were transmitted via
Bluetooth wireless connection in real time to an Android phone which stored the signal on an SD-card for
further processing. Figure 1 shows the jaw motion sensor (Figure 1(a)) and the Automatic Ingestion
Monitor module (Figure 1(b)); Android phone is not shown in the picture.
Figure 1. Wearable sensor system for monitoring jaw movements: (a) jaw motion sensor, (b)
wireless module.
2.8 Sensor data and signal processing
Collected jaw motion sensor signals, further denoted were processed in the following manner.
First the signals were demeaned by subtracting the average computed over the duration of the experiment
from every data point. Next, the entire signals were de-noised using wavelet transform. De-noising
attenuates small variations in the signal due to noise on the power lines (variation in voltage due to
wireless transmission that leaks into the sensor signal). A bi-orthogonal wavelet transform (a 1-D
symmetric wavelet with symmetric boundary treatment, using a lifting implementation) with 4 vanishing
moments was used for de-noising. The sensor signal was decomposed using Discrete Wavelet Transform
(DWT) technique, the wavelet coefficients below a threshold were discarded and the de-noised sensor
signal was recovered using Inverse Discrete Wavelet Transform. This threshold was found as described
further in the manuscript. Figure 2 shows a segment of the jaw motion sensor signal before and after de-
noising.
Figure 2. Jaw motion sensor signal before (a) and after de-noising (b) using the bi-orthogonal
wavelet transform. Amplitude is in the Analog to Digital convertor (ADC) units.
2.9 Sucking count, sucking rate and error computation for sensor signal
After de-noising, the sensor signals were divided into epochs of 10s each, denoted as ,
with samples within each epoch, where , =1000Hz is the sampling frequency and =10s
is the epoch size. These epochs were time-synchronous with the 10s epochs used during the signal
annotation process. For each epoch, sucking count and sucking rate were computed from the sensor signal
by the algorithm shown in Figure 3.
Figure 3. Algorithm for computation of per epoch computer-predicted sucking count and
sucking rate for n-th epoch. is in sucks per seconds.
For any epoch n, computer-predicted counts were compared to human-annotated counts
to evaluate the performance of the method. For each infant, per epoch mean error
was computed as follow:
(1)
where is the total number of epochs and represents the infant. The mean per epoch sucking rate error
has exactly the same value as
, as per epoch sucking rate is the average of sucking
count over a period of 10s. Since these errors are computed on a per-epoch basis, they illustrate accuracy
of the method in evaluating changes in the suck count and sucking rate over time.
The cumulative error over a meal for infant k was computed as the difference of the sum of all epochs:
(2)
Similarly, the cumulative sucking rate error has the same numeric value as . Cumulative
errors show performance of the method averaged over the meal duration, where under and over
predictions on individual epochs may compensate each other. Thus, cumulative error is expected to be
lower than per-epoch error. Per epoch mean error for sucking rate and sucking count will have the same
value whereas the cumulative error (for both sucking rate and sucking count) over an entire meal will
have a different value than per epoch error.
2.10 Parameter determination and validation
In order to find a generalized threshold for the de-noising process and to evaluate accuracy of the
method, a leave one out cross validation scheme was used. For de-noising, this procedure found a scaling
factor α that was used to compute the de-noising threshold as a function of the jaw sensor signal’s
Set the variable (number of mean crossings) to 0.
For th sample in th epoch, increment by 1,
if
and
or
where =1,…, 2, and
is the mean amplitude of the epoch .
Compute per epoch sucking count as .
Compute per epoch sucking rate /10.
amplitude: , where is the standard deviation of the signal. The scaling factor α
for infant k was found by with-holding infant k from the dataset and performing a grid search for a value
of on the dataset from the remaining 9 infants (training set). The value of α which resulted in
the minimal absolute average
on the training set, was used to validate performance of the method
on the withheld (validation) data from infant k by computing corresponding
and .
Absolute value was used to account for signed error. This process was repeated such that each infant was
used for the validation set only once.
Measures of population-wide performance were then computed as averages of absolute values of
and over validation results from all 10 infants:
2.11 Intra-class correlation (ICC) and statistical analysis
Two Intra-class correlation (ICC) analyses were conducted on clusters (entered by each observer)
across ten videos (6 breastfed and 4 bottle-fed) based on a 2-way mixed model with observers fixed and
subjects random to examine absolute agreement. First ICC analysis was performed between two human
raters to determine their reliability. Second ICC analysis for the raters (averaged together) and the jaw
sensor algorithm was conducted to compare the reliability of sensor signal against human raters.
Infant weight-for-age, length-for-age, and body mass index-for-age z-scores were calculated using the
World Health Organization (WHO) Anthro software (version 3.2.2, January 2011), which is based on
international growth charts of healthy infants growing under optimal conditions [24]. Since factors such
as age and the feeding mode (breast-fed vs. bottle-fed) can impact the feeding behavior, a comparison was
made between the annotated and computer-predicted sucking rate (sucking counts) for different feeding
modes, gender, BMI-for-age, Weight-for-age and Length-for-age z-scores and age, % fat, total fat mass,
and total fat free mass to examine whether there was any effect of these variables on the accuracy of the
computer-predicted counts. Two tailed, two sample t-test was used to determine the effect of feeding
mode on the suck count and sucking rate errors. A Passing-Bablok regression [25] was used to determine
the effects of BMI (BMI-for-age z-scores), body weight (Weight-for-age z-scores), body length (Length-
for-age z-scores), age, % body fat, total fat mass and total fat free mass, on the performance of the
proposed technique. Passing-Bablok regression is not sensitive towards outliers and assumes that
measurement errors from both variables have same distribution. All analyses were conducted using
MATLAB ® (Mathworks Inc.).
3. RESULTS
The ICC analysis conducted on a total of 692 clusters resulted in an average correlation coefficient
value of 0.98, 95% CI [0.98, 0.99], for sucking counts between the two raters. For sucking count, the
average correlation coefficient value for absolute agreement for the sucking algorithm and the average of
the two raters for 692 clusters was 0.86, 95% CI [0.83,0.88]. ICC results for sucking rate were same as
above, since per epoch sucking rate is the average of sucking count over a period of 10s.
(3)
(4)
For per epoch sucking count/rate estimation, the sensor-based method resulted in population-wide
mean absolute errors
18.52% (SD +/-11.17%) and
18.85% (SD +/-10.45%)(Cumulative
sucking count/rate error). Per-infant errors are summarized in Table 2. Figure 4 shows an example of the
annotated and the computer-predicted sucking rate for an infant over the period of an entire experiment.
A comparison was done between bottle-fed and breast-fed infants in-terms of per epoch mean error for
suck count/rate. The proposed technique achieved a per epoch suck count/rate mean absolute error of
26.62% (SD +/-14.16%) for bottle-fed infants and for breast-fed infants the algorithm was able to achieve
per epoch mean absolute error of 13.14% (SD +/-4.14%). A two tailed, two sample t-test was used to
show that the mean errors for both groups (bottle-fed and bottle fed) are significantly different (p=0.02),
for 95% CI. The method was able to achieve per epoch suck count/rate mean absolute error of 25.17%
(SD +/-12.61%) and 11.89% (SD +/-3.40%) for male and female infants respectively. Results of two
tailed, two sample t-test showed that the errors are not dependent on the gender of the infant (p = 0.22),
for 95% CI.
Figure 5(a) shows the results of Passing-Bablok regression analysis of
vs. BMI-for-age z-
scores. Figure 5(b), Figure 5(c) and Figure 5(d) show Passing-Bablok regression analysis
between
and corresponding Weight-for-age z-scores, Length-for-age Z-score and age (in
weeks), respectfully. Figure 6 show the regression results of
vs. % body fat (Figure 6(a)), total
fat mass (Figure 6(b)) and total fat free mass (Figure 6(c)). All of these tests indicate independence of
.
Table 2 Result: Comparison between annotated and computer-predicted sucking rate and sucking count with error.
Infant
Feeding
mode
Human-
counted
sucks
Computer-
predicted
sucks
Human-
estimated
sucking
rate
Computer-
estimated sucking
rate
Per-epoch mean
count/rate error
Per-meal mean
count/rate
error
1
Breast-fed
354
250
0.77
0.54
8.88%
29.28%
2
Breast-fed
774
701
0.87
0.79
-10.07%
9.43%
3
Breast-fed
859
946
0.98
1.07
-9.55%
-10.07%
4
Breast-fed
979
748
1.24
0.95
18.53%
23.60%
5
Breast-fed
329
263
1.03
0.82
16.65%
20.09%
6
Breast-fed
621
696
0.71
0.79
-15.15%
-8.50%
7
Bottle-fed
416
577
0.72
0.99
-40.95%
-38.70%
8
Bottle-fed
662
833
1.00
1.26
-36.61%
-25.85%
9
Bottle-fed
1023
1104
1.31
1.41
-14.61%
-7.87%
10
Bottle-fed
603
694
1.06
1.22
-14.29%
-15.10%
Mean:
-9.72%
-2.37%
STD:
20.03%
22.31%
Mean (abs value):
18.52%
18.85%
STD (abs value):
11.17%
10.45%
* Sucking rate is in the units of sucks per s, represents infant’s number in the dataset.
Figure 4. A comparison of annotated sucking rate vs. computer-predicted sucking rate computed
over 10s epochs for the entire experiment (infant 9). Epochs where the both values are
zeros are the epochs which involved no intake or where the raters were not able to score
the signals.
Figure 5. (a) Passing-Bablok regression analysis for
and BMI-for-ageZ-score.
Regression line equation y = -0.01 - 0.15x; 95 % CI, for intercept -0.06 to 0.06 and for
slope -0.34 to 0.23. (b) Passing-Bablok regression analysis for
and Weight-
for-age Z-score. Regression line equation y = 0.02 - 0.15x; 95 % CI, for intercept -0.02 to
0.03 and for slope -0.32 to 0.28. (c) Passing-Bablok regression analysis for
and Length-for-age Z-score. Regression line equation y = -0.06- 0.18x; 95 % CI, for
intercept -0.05 to 0.46 and for slope -0.1733 to 1.25. (d) Passing-Bablok regression
010 20 30 40 50 60 70 80
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Sucking rate
Epoch number (10 seconds each)
Annotated sucking rate
Predicted sucking rate
analysis for
and age (in weeks). Regression line equation y = 0.09 - 0.01x; 95
% CI, for intercept -0.81 to 2.23 and for slope -0.12 to 0.05.
4. DISCUSSION
Research shows that feeding behavior and weight gain during first 6 months of infancy is associated
with risk for obesity later in life. Therefore it is important to accurately and objectively monitor feeding
behavior in infants. This study introduced a new approach for objective monitoring of feeding behavior in
infants using a jaw motion sensor and the sucking count signal processing method. In the experiment, the
new sensor system was tested on a population consisting of both breast-fed and bottle-fed infants to show
its feasibility for both groups over a wide range of age and BMI status. Results presented show a
comparison of sucking counts and sucking rate between the human raters and what was predicted by the
proposed method. In general, these results suggest that although there was a good correlation between the
sucking count detected by human raters and that estimated by the jaw sensor data, further studies are
needed to examine reasons for, and methods to reduce, the absolute error rate of this method.
Reliability of the manual scores by the human raters was verified by conducting ICC statistics.
Results from the ICC analysis shows that there is a high inter-rater reliability (0.98) for sucking count
computation between the two raters. A second ICC statistic computed for the average sucking count of
two raters (combined) and the computer-predicted sucking count/rate computed by the algorithm from the
sensor signal also revealed a high value of ICC (0.86) shows that there is a strong agreement between the
values annotated by the human raters and the computer-predicted values. This agreement is critical for
increasing the confidence in the algorithm results.
Figure 6. (a) Passing-Bablok regression analysis for
and % body fat. Regression line
equation y = -0.03- 0.01x; 95 % CI, for intercept -1.25 to 1.17 and for slope -0.05 to 0.05
(b) Passing-Bablok regression analysis for
and Total fat mass (ing).
Regression line equation y = -0.65 +0.39x; 95 % CI, for intercept -2.57 to 0.38 and for
slope -0.23 to 1.60 (c) Passing-Bablok regression analysis for
and Total fat
free mass (in g). Regression line equation y = 1.19-0.24x; 95 % CI, for intercept –2.67 to
2.99 and for slope –0.61 to 0.52.
For algorithm development there were two possibilities for finding an optimum threshold and
parameter selection. One possibility was to optimize per epoch error for sucking count/rate and the other
possibility was to optimize parameters for cumulative error for sucking count/rate. Per epoch sucking
count/rate provides a better estimate of the feeding behavior during a meal compared to the cumulative
suck count/rate and therefore it was chosen for parameter optimization. Different results will be obtained
if the parameters were optimized for cumulative suck count/rate. For de-noising threshold was chosen
as a factor of the sensor signal's standard deviation (STD) to account for signal variation during the
experiment. Other possibilities for threshold are signal's maximum or median amplitude.
In terms of per-epoch (18.52 %) and per-meal (18.85 %) average absolute sucking count errors (and
corresponding sucking rate error), the proposed method demonstrated a close match with the human
annotations. High errors and standard deviation indicate the variation of performance of the method
among infants, and further studies are required to identify the source of these variations. Negative values
of the errors for an infant indicate that on average, the proposed algorithm over-estimated the suck counts.
It is possible that this over-estimation is attributable to the presence of non-nutritive sucking periods
where the infant is not actually sucking but just moving his/her lips. Non-nutritive sucking can increase
the computer-predicted sucking count which will induce error. Body movement (motion artifacts) can
also induce false positives in the sensor signal. It is also possible that small sucks were detected by the
sensor but not by the human raters because the chin movement generated by the suck was almost
imperceptible. Addition of other sensor modalities such as monitoring of swallowing may help in
eradicating false positives.
An advantage of the proposed method is its potential to predict the feeding behavior for both feeding
modes (bottle-fed vs. bottle-fed). Results suggest that this technique resulted in a higher per epoch mean
error (as well as standard deviation) for bottle-fed infants compared to breast-fed infants. For performance
comparison, the p-value of 0.02 for the two tailed t-test showed that the performance of this method is
dependent on the feeding mode i.e. the technique performed better for breast-fed infants compared to
bottle-fed infants. Although not examined here, it is possible that differences in infant movement during
the meal may have contributed to greater error among bottle-fed infants, or that bottle-fed infants engage
in more non-nutritive sucking during the meal in comparison to breast-fed infants. The source of this
feeding-mode variation in accuracy of the method needs further investigation. Another possibility is to
analyze the strength/amplitude of the sensor signal for breast-fed and bottle-fed infants separately, and
then implement different thresholds for each group. Results of t-test for gender indicated that the
performance is independent of the infants' gender and the technique works equally well for both genders.
Given that excess body fat or that differences in fat free mass (i.e. muscle) may have contributed to
error in the sensor signal, we examined whether the sensor-estimated sucking count was independent of
various measures of infant body size and composition, such as BMI-for-age, weight-for-age, and length-
for-age z-scores, along with infant total fat mass, total fat free mass, and total % body fat. Results of
Passing-Bablok regression analyses indicates that in this cohort, the performance of this method to
estimate sucking count per epoch was independent of these measures of infant body size and composition.
However, given the small sample size, it would be of interest to replicate these analyses in a larger cohort.
The small sample size is the main limitation of this study. It will be important to replicate this study
in a larger cohort in order to confirm or extend the findings. Another possible limitation is that the sensor
used in these experiments was originally designed to be used by adults and the size may need to be further
reduced for use with infants. However, there was no indication that the sensor size caused any discomfort
for the infants or mothers, and the sensor did not impede bottle or breastfeeding. Nonetheless, a smaller
sensor may help to reduce the overestimation of sucking count found here, and such an adaptation is
technologically possible by incorporating the electronics of the sensor system into the infant’s clothing. It
is also important to acknowledge that the experiments of this study were performed in laboratory
conditions, and free living tests will be required to evaluate performance of the proposed technique over
extended periods of time under realistic conditions of daily living.
Future studies might examine whether meal and sucking burst boundaries can be automatically
recognized. In the current study, meals were marked by human raters. Relative strength of the sucking
signal and swallowing rate may provide indicators suitable for differentiation of food-related sucking with
sucking on thumbs, pacifiers and vocalizing and should be explored. The strength of the sucking signal
may also be another measure or index to examine in relation to obesity.
Overall, the results of this study suggest that monitoring of jaw motion may potentially be a
promising foundation for monitoring of infant feeding behavior.
5. CONCLUSIONS
This paper evaluated the technical feasibility of using a jaw motion sensor for accurate and objective
monitoring of feeding behavior in bottle-fed and breast-fed infants. The signals captured by the sensor
were processed to estimate sucking counts and sucking rate. The computed counts were compared with
results of human annotation of the same experiments. The mean errors and ICC statistics showed a close,
and acceptable, agreement between the human raters and the proposed methodology. Statistical analysis
of the results suggests that the performance of the proposed method is independent of the factors such as
gender, BMI, length, weight and age of the infants but it did differ by feeding mode. Further study with a
larger population (sample size) is needed to more rigorously examine the statistical significance of these
results. This ensures the applicability of the proposed method to a wider infant population.
ACKNOWLEDGEMENT
This work was supported by a pilot grant from Nutritional Obesity Research Center at the University of
Alabama, Birmingham (UAB). Authors would like to acknowledge the efforts of infants and mothers who
participated in the study as well as students who took part in conducting the experiment and annotating
the data.
The authors have no potential conflicts of interest.
REFERENCE
[1] Mihrshahi S, Battistutta D, Magarey A, Daniels LA. Determinants of rapid weight gain during
infancy: baseline results from the NOURISH randomised controlled trial. BMC Pediatr 2011;11:99.
[2] Gillman MW, Rifas-Shiman SL, Camargo, Jr CA, et al. Risk of overweight among adolescents who
were breastfed as infants. JAMA 2001;285:2461–7.
[3] Dubois L, Girard M. Early determinants of overweight at 4.5 years in a population-based
longitudinal study. Int J Obes 2006;30:610–7.
[4] Stettler N, Zemel BS, Kumanyika S, Stallings VA. Infant Weight Gain and Childhood Overweight
Status in a Multicenter, Cohort Study. Pediatrics 2002;109:194–9.
[5] Taveras EM, Rifas-Shiman SL, Sherry B, et al. Crossing growth percentiles in infancy and risk of
obesity in childhood. Arch Pediatr Adolesc Med 2011;165:993–8.
[6] Ekelund U, Ong K, Linné Y, Neovius M, Brage S, Dunger DB, et al. Upward weight percentile
crossing in infancy and early childhood independently predicts fat mass in young adults: the
Stockholm Weight Development Study (SWEDES). Am J Clin Nutr 2006;83:324–30.
[7] Eriksson M, Tynelius P, Rasmussen F. Associations of birthweight and infant growth with body
composition at age 15 – the COMPASS study. Paediatr Perinat Epidemiol 2008;22:379–88.
[8] Oyama M, Saito T, Nakamura K. Rapid weight gain in early infancy is associated with adult body
fat percentage in young women. Environ Health Prev Med 2010;15:381–5.
[9] Wright P, Fawcett J, Crow R. The development of differences in the feeding behaviour of bottle
and breast fed human infants from birth to two months. Behav Processes 1980;5:1–20.
[10] Taki M, Mizuno K, Murase M, Nishida Y, Itabashi K, Mukai Y. Maturational changes in the
feeding behaviour of infants – a comparison between breast-feeding and bottle-feeding. Acta
Pædiatrica 2010;99:61–7.
[11] F W, Mw W, Cn M, Mj R, Jd B. An ultrasonographic study of the orgenisation of sucking and
swallowing in newborn infants. Pediatr Res 1984;18:806–806.
[12] Woolridge MW, How TV, Drewett RF, Rolfe P, Baum JD. The continuous measurement of milk
intake at a feed in breast-fed babies. Early Hum Dev 1982;6:365–73.
[13] Riordan JM, Koehn M. Reliability and validity testing of three breastfeeding assessment tools. J
Obstet Gynecol Neonatal Nurs JOGNN NAACOG 1997;26:181–7.
[14] Da Costa SP, van der Schans CP, Boelema SR, van der Meij E, Boerman MA, Bos AF. Sucking
patterns in fullterm infants between birth and 10 weeks of age. Infant Behav Dev 2010;33:61–7.
[15] Amft O. A wearable earpad sensor for chewing monitoring. 2010 IEEE Sens., 2010, p. 222–7.
[16] Sazonov E, Fontana JM. A Sensor System for Automatic Detection of Food Intake Through Non-
Invasive Monitoring of Chewing. IEEE Sens J 2012;12:1340 –1348.
[17] Fontana JM, Sazonov ES. A robust classification scheme for detection of food intake through non-
invasive monitoring of chewing. 2012 Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBC, 2012, p.
4891 –4894.
[18] M. Fontana J, Farooq M, Sazonov E. Estimation of Feature Importance for Food Intake Detection
Based on Random Forests Classification, Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBC, 2013,
p.6756 - 6759
[19] Sazonov E, Schuckers S, Lopez-Meyer P, Makeyev O, Sazonova N, Melanson EL, et al. Non-
invasive monitoring of chewing and swallowing for objective quantification of ingestive behavior.
Physiol Meas 2008;29:525–41.
[20] "WHO | Use of World Health Organization and CDC Growth Charts for Children Aged 0-59
Months in the United States"[Online]. Available:
http://www.cdc.gov/mmwr/preview/mmwrhtml/rr5909a1.htm [Accessed: 19-Dec-2013].
[21] Fomon SJ, Haschke F, Ziegler EE, Nelson SE. Body composition of reference children from birth
to age 10 years. Am J Clin Nutr 1982;35:1169–75.
[22] Hernandez-Reif M, Field T, Diego M. Differential sucking by neonates of depressed versus non-
depressed mothers. Infant Behav Dev 2004;27:465–76.
[23] Hernandez-Reif M, Field T, Del Pino N, Diego M. Less exploring by mouth occurs in newborns of
depressed mothers. Infant Ment Health J 2000;21:204–10.
[24] "WHO | The WHO Multicentre Growth Reference Study (MGRS): Rationale, planning, and
implementation."[Online]. Available: http://www.who.int/childgrowth/mgrs/fnu/en/[Accessed: 19-
Dec-2013].
[25] Passing H, Bablok. A new biometrical procedure for testing the equality of measurements from two
different analytical methods. Application of linear regression procedures for method comparison
studies in clinical chemistry, Part I. J Clin Chem Clin Biochem Z Für Klin Chem Klin Biochem
1983;21:709–20.