Sunteți pe pagina 1din 5

2013 12th International Conference on Machine Learning and Applications

A Comparative Study of Food Intake Detection


Using Artificial Neural Network and Support
Vector Machine
Muhammad Farooq, Juan M. Fontana, Member, IEEE, Akua F. Boateng, Megan A. McCrory and
Edward Sazonov*, Senior Member, IEEE

Abstract— In Machine Learning applications, the selection of associated with ED and obesity.
the classification algorithm depends on the problem at hand.
This paper provides a comparison of the performance of the
Current MIB methods include food frequency
Support Vector Machine (SVM) and the Artificial Neural questionnaires, food records and random 24-hour dietary
Network (ANN) for food intake detection. A combination of recalls. All these methods require active participation of
time domain (TD) and frequency domain (FD) features, subjects and are prone to errors due to various reasons,
extracted from signals captured using a jaw motion sensor, including incorrect reporting of food consumed and failure
were used to train both types of classifiers. Data were collected to report certain foods [5]. Therefore, it is important to
from 12 subjects in free-living for a period of 24-hrs under investigate automatic and more accurate methods for
unrestricted conditions. ANN with a different number of objective monitoring of food intake in free living conditions.
hidden layer neurons and SVMs with different kernels were
trained using a leave one out cross validation scheme. ANN Innovative solutions based on wearable sensor systems
achieved an average accuracy of 86.86 ± 6.5 % whereas SVM have been developed for automatic, non-invasive and more
(with linear kernel) achieved an average classification accuracy accurate food intake detection and characterization. These
of 81.93 ± 9.22 %. Data collected from an independent subject methods use features related to different stages of the food
in a separate study were used to evaluate the performance of consumption process (hand gestures, bites, chewing and/or
these classifiers in-terms of the number of meals detected per swallowing) together with signal processing and pattern
day resulting in an accuracy of 72.72% for ANN and 63.63% recognition algorithms for MIB. In [6] and [7], a microphone
for SVM. The results suggest that ANN may perform better placed in the ear canal was used to capture sounds generated
than SVM for this specific problem. during chewing and/or swallowing during food intake. In
[6], Hidden Markov classifiers achieved 83% detection
Keywords—Food intake detection, Neural Net, SVM, accuracy for food intake. In [7], a clustering approach was
chewing, eating disorder, wearable sensors. used to achieve a recognition rate of about 86%. In [8], an
Artificial Neural Network (ANN) achieved a classification
I. INTRODUCTION accuracy of around 84% from chewing sounds captured
Monitoring of Ingestive Behavior (MIB) is particularly through a microphone placed in the ear canal. Swallowing
important in populations suffering from eating sounds captured through a throat microphone and jaw
disorders(ED) and obesity. Eating disorders cause the person motion captured through a piezoelectric strain gauge sensor
to consume either insufficient or excessive amounts of food were used in [9] and [10] respectively; to train Support
which results in abnormal eating behavior [1]. Some of the Vector Machines (SVM) for food intake detection. In [10],
most common eating disorders are anorexia nervosa, bulimia individual models achieved an average classification
nervosa and binge eating disorder with lifetime prevalence accuracy of 81%. In [9] SVM classifiers resulted in average
ranging from 0.6 to 4.5% in USA [2]. Obesity is the detection accuracy of 84.7% for swallowing events. In [11],
condition of excessive body fat accumulation due to an the supervised method, SVM, and the unsupervised method,
imbalance between energy intake and energy expenditure. k-means, are compared for food intake detection with the
Only in the USA, 35.3% of the population is suffering from results indicating better performance of the unsupervised
obesity [3]. Obesity is the 5th major cause of death method. These experiments were performed in a controlled
worldwide and about 2.8 million people die each year from a laboratory environment. The performance of these pattern
disease related to obesity [4]. Therefore, obtaining an recognition algorithms would be hindered while testing out
accurate and objective monitoring of food intake is of the lab due to the presence of artifact and activities that
important for identification and diagnosis of eating patterns cannot be replicated in the laboratory setup. Therefore, it is
extremely important to measure the classifier performance
The project described was supported by Grant Number R21DK085462 from
under unrestricted conditions of community living, such as
the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) the tests conducted in this paper.
and Purdue Ingestive Behavior Research Center (IBRC). The content is solely
the responsibility of the authors and does not necessarily represent the official Performance of a classifier is highly dependent on the
views of the NIDDK or the NIH or the Purdue IBRC. problem being addressed [12]. In this study, we compared
Muhammad Farooq (mfarooq@crimson.ua.edu), Juan M. Fontana the performance of two different classifiers for food intake
(jmfontana@bama.ua.edu) and Edward Sazonov (esazonov@eng.ua.edu) are
with the Department of Electrical and Computer Engineering, University of
detection: ANN trained with a different number of hidden
Alabama, Tuscaloosa, AL 35487 USA. Akua Boateng (aboaten@purdue.edu) layer neurons and SVM trained with different kernels. Both
and Megan A. McCrory (mmccror@purdue.edu) are with the Department of classifiers were trained with features extracted from jaw
Nutrition Science and the Purdue IBRC. Dr. McCrory also has a joint motion sensor signals captured through a wearable sensor
appointment with Department of Psychological Sciences, Purdue University,
West Lafayette, IN 47907. system (Automatic Ingestion Monitor, AIM). A total of 12
*Corresponding author (phone: 205-348-1981) subjects were asked to wear this sensor system for 24-hrs

978-0-7695-5144-9/13 $31.00
$26.00 © 2013 IEEE 153
DOI 10.1109/ICMLA.2013.33
under free living conditions and resulting signals were used
to train the classifiers. Data collected from two days from
one subject at a separate independent location were used to
test these classifiers. Results indicate that ANN performed
better than SVM for this specific problem.

II. METHODS
A. Data Collection
Data were collected at two independent locations for
training and testing of the classifier (Training data set at the
University of Alabama and Test data set at Purdue
University). Training data were collected from a total of 12
subjects (six males, six females) in this study (at the Fig. 1. Subject wearing the wearable sensor system
University of Alabama). The average age was 26.7 y (SD ±
3.7) and the average body mass index (BMI) was 24.39 B. Signal Preprocessing and Feature Extraction
kg/m2 (SD ± 3.81). The test data set was part of an
independent study performed at Purdue University; where The jaw motion signal, JM (t), and the accelerometer
data were collected from a subject (female) for two days signals, ACCX(t), ACCY(t), and ACCZ(t), were high-pass
(~48hrs) using the AIM. Subject kept a brief time record of filtered (0.1 Hz cutoff frequency) to remove the DC
their eating occasions on each day. Selected population at component. JM (t),ACCX(t), ACCY(t), and ACCZ(t) were then
both locations did not present any medical condition that normalized to compensate for variations in signal amplitude
would hinder normal eating behaviors. The Internal Review between subjects. The hand gesture signal, HG(t), was
Boards at The University of Alabama and at Purdue normalized with respect to its maximum value. HtM gestures
University approved these studies. Subjects had to read and shorter than 0.25s and longer than 7.5s were removed from
sign an informed consent form before the start of the HG(t) as it was assumed that they did not belong to food
experiment. intake activity.

Each subject wore the sensor system for a period of 24-hrs


under free living conditions while performing their daily
routines without restrictions. A picture of a subject wearing
the sensor system is shown in Fig. 1. The wireless sensor
module consisted of three sensors:
a) A jaw motion sensor to detect characteristic motion of
the jaw during chewing [13][14]. This sensor was attached
directly below the ear using medical adhesive.
b) A hand gesture sensor to detect hand-to-mouth gestures
(HtM) associated to bites. It consisted of a RF transmitter
worn on the inner side of the dominant arm and a RF receiver
in the wireless module operating in RFID frequency band of
125 KHz.
c) A tri-axial accelerometer located in the wireless module
to detect body acceleration.
Sensor signals were captured by the wireless module at a
sampling frequency of 1 kHz and were transmitted via
Bluetooth in near real time to an Android phone which
worked as a data logger. An example of the signals collected Fig. 2. Example of the signals collected after a subject wore the wearable
after 24hrs is shown in Fig. 2. sensor system for 24hrs.

For the training set, a push button was used for self- The jaw motion sensor signals were divided into non-
reporting food intake by the subjects. Subjects pressed and overlapping segments of 30s called epochs. This epoch size
held the button during chewing to mark food intake. helps to detect small food intake events such as snacking.
Subjects were also asked to keep a log of their daily Time and frequency domain features were computed for each
activities as well as the start and end times of each meal. epoch of the jaw motion sensor signal (Table 1). For
Push button signals combined with the activity logs were computation of frequency domain features, the spectrum of
used as a gold standard for assigning class labels to food jaw motion sensor signal was divided into three different
intake data. For the test set, the class labels were assigned frequency ranges belonging to different activities (1.25-2.5
using only log data as subject was asked to keep a log of her Hz for chewing, 2.5-10 Hz for walking, and 100-300 Hz for
activities, but was not required to use the push button. The talking [14]). This information could potentially be used by
test data were used to determine number of meal episodes the classifier for discrimination between food intake and no
per day. food intake epochs. Features listed in Table 1 constituted
feature fi set for ith epoch.

154
Each feature vector fi was associated with a class label ci E. Leave One Out Cross Validation and Testing
= {'no food intake'; 'food intake'} for the classification task. A leave one out (12 fold) cross validation scheme was
A class label belonged to 'food intake' (ci =1) if at least 10s of used to train ANN and SVM classifiers. Data from 11
the self-report signal within the i-th epoch reported food subjects was randomly divided into training (80% of the
intake; otherwise, it belonged to 'no food intake' (ci=-1). data) to train the classifiers and validation (20% of the data)
C. Artificial Neural Network (ANN) for validating the classifiers. Validation data was used for
parameters tuning for both classifiers. Testing of the
Artificial neural network (ANN) is a supervised learning classifiers was performed with data from the subject that was
technique that has shown excellent results for a number of left out (12th subject). This procedure was repeated 12 times
pattern recognition and classification problems [15]. Neural so each subject could be used as the test subject. The
networks are robust and flexible and can analyze complex classification accuracy for each classifier was computed and
patterns with less formal training and can handle noisy data. averaged to get the final results. Per epoch classification
A three layered (input layer, hidden layer and output layer) accuracy was used to evaluate the performance of the
feed-forward neural network with back-propagation training classifiers. Classification accuracy was defined as the
algorithm was used for food intake classification. The average between Precision and Sensitivity (Recall) in order
proposed ANN architecture consisted of an input layer with to account for the high number of true negatives that are
38 input neurons (one for each predictor), one hidden layer typical in the monitoring of food intake over long periods.
with N hidden neurons and one output layer with one output
neuron. N ranged from 1 to 10 neurons to obtain the 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙) /2 (1)
optimum ANN classifier. 𝑇+ (2)
𝑃𝑟𝑒𝑐𝑖𝑠𝑜𝑛 =
TABLE 1. TIME AND FREQUENCY DOMAIN FEATURES EXTRACTED FROM 𝑇+ + 𝐹+
𝑇+ (3)
EACH EPOCH OF THE JAW MOTION SIGNAL
𝑅𝑒𝑐𝑎𝑙𝑙 =
# Description # Description 𝑇+ + 𝐹−
Energy spectrum in chewing
1 Mean Absolute Value (MAV) 20
range (chew_ene) Where 𝑇+ was the number of food intake epochs correctly
Entropy of spectrum chewing
2 Root Mean Squared (RMS) 21
range (chew_ene)
classified by the classifier as food intake, 𝐹+ was the number
3 Maximum value (Max) 22 chew_ene / spectr_ene of no food intake epochs incorrectly classified as food
Energy spectrum in walking intake, and 𝐹− was the number of food intake epochs
4 Median value (Med) 23 incorrectly classified as no food intake. Since the number of
range (walk_ene)
5 MAV / RMS 24
Entropy of spectrum walking true negatives 𝑇− (periods of no food intake correctly
range (chew_ene) classified as no food intake) were not involved in the
6 Max / RMS 25 walk_ene / spectr_ene computation of accuracy, the rate of detection of no food
Energy spectrum in talking intake did not affect the overall accuracy of food intake
7 MAV / Max 26
range (chew_ene)
Entropy of spectrum talking
detection. Since in the test data the push button signal is
8 Med / RMS 27 missing, it was not possible to the use same type of accuracy
range (chew_ene)
9 Signal entropy (Entr) 28 talk_ene / spectr_ene measure for test data as is used for training data.
10 Num. of zero crossings (ZC) 29 chew_ene / walk_ene Performance of the best classifiers on the test data was
11 Mean time between ZC 30 chew_entr / walk_entr evaluated in terms of the number of meals correctly
12 Num. of peaks (NP) 31 chew_ene / talk_ene
detected.
13 Average range 32 chew_entr / talk_entr Both training and test data sets followed the same pre-
14 Mean time between peaks 33 walk_ene / talk_ene processing and feature extraction procedures. For assessment
15 NP/ZC 34 walk_entr / talk_entr of patterns of food intake in free living conditions, it is
16 ZC/NP 35 Fractal dimension (fractal_d) important to determine the number of meals consumed per
Peak frequency in chewing day and their duration rather than individual food intake
17 Wavelength 36
range (maxf_chew) epochs. To assess the ability of the trained classifiers to
Peak frequency in walking detect the number of meal episodes consumed per day, two
18 Num. slope sign changes 37
range (maxf_walk) days data from an independent subject (data from Purdue
Energy of frequency spectrum Peak frequency in talking
19
(spectr_ene)
38
range (maxf_talk) University) was presented to the best ANN and SVM
classifiers. When food intake activity is detected with a gap
of less than five minutes, it was considered to be part of the
D. Support Vector Machine (SVM) same meal. The number of meals predicted by the classifiers
Support vector machine is a machine learning was compared to the daily activity logs to evaluate the
technique that maps the input data from a lower dimension performance of the classifiers.
to a higher dimensional feature space where the data can be
linearly separated by a hyper-plane (using different kernel III. RESULTS
functions). This study investigated the use of four different Tables 2 and 3 show the cross validation results of the
kernel functions for food intake classification: Linear ANN and the SVM classifiers. These results were obtained
Kernel, Radial Basis Function (RBF), Polynomial (3rd by repeating 12 fold cross validation 10 times for ANN and
degree) and Gaussian Radial Basis Function. SVM was SVM and final results were obtained by averaging. Table 2
implemented using the LibSVM software package [16]. shows the results for ANN with different numbers of hidden
layer neurons. Best results were obtained for ANN with 5

155
neurons in the hidden layer resulting in a Precision of 87.59 Signals used to train the classifiers were captured from a
± 7.93 %, a Sensitivity of 86.13 ± 7.28% and an average jaw motion sensor modality integrated into a wearable
accuracy of 86.86 ± 6.5%. Average accuracy remains nearly sensor system. Subjects wore the AIM in free living for a
the same for higher numbers of neurons in the hidden layer. period of 24hrs without imposing any restrictions on daily
activities. The resulting dataset included high intra and inter-
Table 3 shows the results of food intake detection using subject variability. Different activities (i.e. walking, talking,
SVM with different kernel functions. The best results were eating, sleeping, resting, etc.) performed by the subjects over
obtained for SVM with linear kernel with a Precision a period of 24-hrs results in intra-subject variability. Inter-
of83.76 ± 11.55%, a Sensitivity of 80.10 ± 10.11% and an subject variability was achieved from the 12 subjects
average accuracy of 81.93 ± 9.22%. For this application, the participating in the data collection. Both ANN and SVM
results indicated that ANN on average has better classifiers were able to incorporate such variability into
classification accuracy than SVM. subject-independent classifiers that achieved an acceptable
Best classifiers from both types of classifiers were used accuracy for food intake detection with 30 s time resolution.
on the test data set to determine the number of meals. In two TABLE 4. CLASSIFIER PERFORMANCE OVER THE TEST DATA SET
days, the test subject consumed a total of eleven meals out of Day Meal Meal contents ANN SVM
which ANN was able to correctly identify eight meals Episode Detected Detected
whereas only one non-meal episode was classified as meal 1 Fiber plus mixed nuts bar Yes Yes
(false positive). On the other hand SVM correctly detected 2 Fish, water and fries Yes Yes
seven meals and missed four meals but had a higher number 3 Coffee and brownie* Yes No
Day
4 Pesto Polenta, Water, Cherry Yes Yes
of false positives (three). Results on the test data are given in 1
tomatoes
Table 4. 5 Apple cider, water, brownie* No No
6 Coffee * No No
TABLE 2.VALIDATION RESULTS OF ANN WITH HIDDEN NEURONS BETWEEN
1 AND 10 (MEAN ± S.D.*) 1 Granola Bar, water Yes Yes
2 Taipei shrimp, fried rice, Yes Yes
# of hidden Precision (%) Sensitivity (%) Accuracy (%) water and water melon
neurons Day 3 Triscuites crackers, humus Yes Yes
1 83.25 ± 11.11 82.97 ± 13.38 83.11 ± 9.27 2 (roasted pine nut)
2 85.49 ± 11.25 84.18 ± 8.42 84.83 ± 7.59 4 Black coffee* No No
5 Taipei shrimp, fried rice, Yes Yes
3 85.19 ± 8.95 84.92 ± 8.9 85.06 ± 7.55 water
4 85.27 ± 11.28 86.73 ± 8.25 86.00 ± 6.67 * Mostly liquid consumption and will probably be missed by the classifier
5 87.59 ± 7.93 86.13 ± 7.28 86.86 ± 6.5 due to lack of chewing.
6 86.46 ± 9.11 84.4 ± 8.81 85.43 ± 7.86 Results of 12 fold cross validation showed that ANN
7 86.72 ± 9.29 86.55 ± 4.21 86.63 ± 6.2 performed better than SVM. Best results were obtained for
8 87.41 ± 10.98 84.81 ± 6.33 86.11 ± 6.49 ANN with five hidden neurons. Average classification
9 88.24 ± 8.48 84.9 ± 8.98 86.57 ± 7.42 accuracy increases when number of hidden neurons is
increased from 1 to 5. There is a slight decrease in average
10 86.59 ± 9.87 86.18 ± 11.71 86.38 ± 8.55
accuracy for hidden neurons above five. Having a small
*S.D. = Standard Deviation
number of neurons in the hidden layer will result in a
TABLE 3.VALIDATION RESULTS OF SVM CLASSIFIER WITH SEVERAL classifier that is not able to classify complex patterns. On the
KERNEL FUNCTION (MEAN ± S.D) other hand having a high number of hidden neurons can
Kernel Precision (%) Sensitivity Accuracy (%) cause over-fitting to the training set. Therefore, it is
function (Recall) (%)
important to find the optimum number of hidden neurons,
Linear 83.76 ± 11.55 80.10 ± 10.11 81.93 ± 9.22
five in this case. Results suggest that selection of an
RBF 86.02 ± 12.23 76.94 ± 9.95 81.48 ± 9.09
appropriate kernel function is important for SVM. For this
Polynomial 86.09 ± 14.15 68.75 ± 19.70 76.92 ± 14.59
specific application, linear and RBF kernels performed
Gaussian 81.93 ± 13.15 77.29 ± 14.63 79.61 ± 11.94
similarly. Best results were achieved for linear kernel
IV. DISCUSSION whereas worst results were obtained for polynomial kernel
function. This indicates that the simplest kernel function was
Automatic, accurate and objective monitoring of food better suited for this application.
intake in the free living condition is important in order to
understand the eating behavior of people. An objective Comparison of the results suggests that, on average,
detection can be obtained using pattern recognition ANN classifiers were able to achieve better classification
algorithms whereas an accurate detection is more accuracy than SVM classifiers for this application. The
challenging and could be obtained by extracting relevant nature of ANN training allows more variability compared to
information from the sensor signals. Since the performance SVM (random initialization of the weights and bias).
of a pattern recognition system is highly dependent on the Another reason can be the presence of complex patterns in
underlying characteristics of the problem, different the data which SVM was not able to separate by the
algorithms may perform differently under the same resulting hyper-plane after training. ANN uses parametric
circumstances. This paper presented a comparative study of classifiers whereas SVM are non-parametric. As ANN
the performance of ANN and SVM classifiers for food suffers from the problem of over-fitting greatly compared to
intake detection. SVM, data from two completely independent subjects were
used to test the best classifiers (ANN with five hidden

156
neurons and Linear SVM) and compare their performance [6] S. Päßler, M. Wolff, and W.-J. Fischer, “Food intake monitoring: an
in-terms of number of meals per day. On test data, ANN acoustical approach to automated food intake activity detection and
classification of consumed food,” Physiol Meas, vol. 33, no. 6, pp.
achieved a better detection rate (eight out of eleven food 1073–1093, 2012.
intake periods (72.72% detection rate)) compared to the [7] O. Amft, “A wearable earpad sensor for chewing monitoring,” in
SVM classifier (seven out of eleven food intake periods 2010 IEEE Sensors, 2010, pp. 222 –227.
(63.63% detection rate)). SVM classifier missed four meals [8] S. Päßler and W.-J. Fischer, “Food Intake Activity Detection Using an
compared to three meal episodes missed by the ANN Artificial Neural Network,” Biomed Tech (Berl), Aug. 2012.
[9] E. Sazonov, O. Makeyev, S. Schuckers, P. Lopez-Meyer, E. L.
classifier. All the three meal episodes missed by the ANN as Melanson, and M. R. Neuman, “Automatic detection of swallowing
well as SVM belonged to the consumption of coffee or other events by acoustical means for applications of monitoring of ingestive
type of liquids. Since liquid consumption does not involve behavior,” IEEE Trans Biomed Eng, vol. 57, no. 3, pp. 626–633, Mar.
chewing therefore in general the chewing sensor will not be 2010.
able to detect liquid consumption. Results shows that the [10] E. Sazonov and J. Fontana, “A Sensor System for Automatic
SVM had a higher number of false positives (three in total, Detection of Food Intake Through Non-Invasive Monitoring of
Chewing,” IEEE Sensors Journal, accepted for publication.
compared to one for ANN).Some of the false positives [11] P. Lopez-Meyer, O. Makeyev, S. Schuckers, E. Melanson, M.
belonged to episodes of exercise or excessive physical Neuman, and E. Sazonov, “Detection of Food Intake from
activity. Having a large number of false positives can be Swallowing Sequences by Supervised and Unsupervised Methods,”
tackled by adding an additional sensor modality such as a Ann Biomed Eng, vol. 38, no. 8, pp. 2766–2774, 2010.
camera. Where image capturing can be triggered by [12] S. S. Lee and J. F. E. IV, Bundling Heterogeneous Classifiers with
Advisor Perceptrons. 1997.
detection of food intake thought the jaw motion sensor and [13] E. Sazonov and J. M. Fontana, “A Sensor System for Automatic
then computer vision techniques can be used to detect actual Detection of Food Intake Through Non-Invasive Monitoring of
presence of food on the table. Based on the results, per Chewing,” IEEE Sens J, vol. 12, no. 5, pp. 1340 –1348, 2012.
epoch accuracies for training set for both classifiers are [14] J. M. Fontana and E. S. Sazonov, “A robust classification scheme for
better compared to the per meal classification accuracies. detection of food intake through non-invasive monitoring of
chewing,” in 2012 Annual International Conference of the IEEE
This is expected as the number observations of epochs Engineering in Medicine and Biology Society (EMBC), 2012, pp.
(training data) are much higher than the number of 4891 –4894.
observations of meal periods (test data). Also for testing set [15] R. Nayak, L. Jain, and B. Ting, “Artificial neural networks in
meal episodes missed mostly belonged to liquid intake such biomedical engineering: a review,” Faculty of Science and
as coffee, which do not involve chewing. Monitoring of Technology, 2001. [Online]. Available:http://eprints.qut.edu.au/1480/.
swallows in addition to monitoring of jaw movement can [Accessed: 28-Jan-2013].
[16] C. Chih-Chung and L. Chih-Jen, “LIBSVM: a library for support
help in detection of liquid intakes. Detection of liquid vectormachines.”[Online].Available:http://www.csie.ntu.edu.tw/~cjlin
intakes will increase the detection accuracy of meal /libsvm.
episodes.
Classifiers trained with combination of time and
frequency domain features were able to detect episodes of
food intake but does not provide any information on what
was consumed. Additional sensors such as a camera with
computer vision algorithms can be used to estimate portion
size and the contents of the meals. An important factor for
any pattern recognition system is the feature set used. A
feature selection algorithm might improve the performance
of this technique. Data used in this experiment was un-
balanced (only 3% of the data belonged to food intake) thus
a data reduction technique would help in achieving more
balanced data for classification algorithms.

REFERENCES
[1] C. G. Fairburn and P. J. Harrison, “Eating disorders,” The Lancet, vol.
361, no. 9355, pp. 407–416, Feb. 2003.
[2] J. I. Hudson, E. Hiripi, H. G. Pope Jr, and R. C. Kessler, “The
prevalence and correlates of eating disorders in the National
Comorbidity Survey Replication,” Biol. Psychiatry, vol. 61, no. 3, pp.
348–358, Feb. 2007.
[3] K. M. Flegal, M. D. Carroll, B. K. Kit, and C. L. Ogden, “Prevalence
of Obesity and Trends in the Distribution of Body Mass Index Among
US Adults, 1999-2010,” JAMA, vol. 307, no. 5, pp. 491–497, Feb.
2012.
[4] “WHO | Obesity and overweight.” [Online]. Available:
http://www.who.int/mediacentre/factsheets/fs311/en/. [Accessed: 21-
Apr-2013].
[5] A. E. Black, G. R. Goldberg, S. A. Jebb, M. B. Livingstone, T. J.
Cole, and A. M. Prentice, “Critical evaluation of energy intake data
using fundamental principles of energy physiology: 2. Evaluating the
results of published surveys,” Eur J Clin Nutr, vol. 45, no. 12, pp.
583–599, Dec. 1991.

157

S-ar putea să vă placă și