PMCCPMCCPMCC

Search tips
Search criteria 

Advanced

 
Logo of nihpaAbout Author manuscriptsSubmit a manuscriptHHS Public Access; Author Manuscript; Accepted for publication in peer reviewed journal;
 
Obesity (Silver Spring). Author manuscript; available in PMC 2010 May 6.
Published in final edited form as:
PMCID: PMC2865137
NIHMSID: NIHMS198608

Toward Objective Monitoring of Ingestive Behavior in Free-living Population

Abstract

Understanding of eating behaviors associated with obesity requires objective and accurate monitoring of food intake patterns. Accurate methods are available for measuring total energy expenditure and its components in free-living populations, but methods for measuring food intake in free-living people are far less accurate and involve self-reporting or subjective monitoring. We suggest that chews and swallows can be used for objective monitoring of ingestive behavior. This hypothesis was verified in a human study involving 20 subjects. Chews and swallows were captured during periods of quiet resting, talking, and meals of varying size. The counts of chews and swallows along with other derived metrics were used to build prediction models for detection of food intake, differentiation between liquids and solids, and for estimation of the mass of ingested food. The proposed prediction models were able to detect periods of food intake with >95% accuracy and a fine time resolution of 30 s, differentiate solid foods from liquids with >91% accuracy, and predict mass of ingested food with >91% accuracy for solids and >83% accuracy for liquids. In earlier publications, we have shown that chews and swallows can be captured by noninvasive sensors that could be developed into a wearable device. Thus, the proposed methodology could lead to the development of an innovative new way of assessing human eating behavior in free-living conditions.

INTRODUCTION

Rates of overweight and obesity are increasing globally. The World Health Organization estimated that there were ~1.6 billion overweight and at least 400 million obese adults worldwide in 2005 and that there will be 2.3 billion overweight and 700 million obese adults worldwide by 2015 (ref. 1). Overweight and obese individuals have an increased risk of developing chronic diseases such as type 2 diabetes, cardiovascular disease, and cancer (25).

Overweight and obesity result from an imbalance between energy intake and energy expenditure, but the etiology of that imbalance and the underlying mechanisms are still incompletely understood. Our physiology, our behavior, and our environment must all be considered in understanding the etiology of obesity. There is a debate about the relative importance of genetic/physiological factors and environmental factors in the etiology of obesity. Clearly there are genetic/physiological contributions to obesity (69) but some weight gain can be attributed to an environment that provides an abundance of inexpensive, highly palatable, and energy dense foods, while requiring only minimal levels of physical activity (1013). Part of our lack of understanding of the etiology of obesity is the fact that most weight gain likely occurs due to very small differences between energy intake and energy expenditure, necessitating very accurate measurements of energy intake and energy expenditure.

A variety of methods are available for accurate and objective measurement of energy expenditure and its components, including doubly labeled water, indirect calorimetry, and accelerometry (1416). These techniques can be used in the laboratory and in free-living populations. Energy and food intake can be accurately monitored in the laboratory by directly measuring consumed food. It is currently not possible to accurately monitor food intake in free-living populations. Several methods have been proposed to measure free-living food intake including observation, weighed food records, estimated records, diet history, food-frequency questionnaires, food recall methods, etc. (17). In a review of 43 studies comparing these methods to indirect measurement using doubly labeled water, the majority suffered from underestimation of energy intake on the order of 0.83 (ratio of intake estimate to energy expenditure) (18).

There is an urgent need for innovative strategies for the accurate assessment of free-living energy intake and monitoring of ingestive behavior in humans. The goal of this study is to explore whether accurate and objective information about ingestive behavior, such as detecting short episodes of food intake, differentiating between liquid and solid foods, and estimating the mass of food intake could be obtained by measuring chews and swallows. As chewing and swallowing events can potentially be captured by noninvasive wearable sensors, this could lead to the first accurate method of assessing free-living food intake. This method may lead to a more objective direct measurement of total energy intake when combined with other methods (e.g., electronic food diaries, food photography, etc.).

METHODS AND PROCEDURES

The models reported in this paper were built using data from a human study described previously (19). A total of 20 volunteers (11 men and 9 women) with an average BMI of 29.0 ± 6.4 participated in four visits, each of which consisted of: (i) a 20-min resting period; (ii) a meal period of unlimited time; (iii) a second 20-min resting period. Two fixed sizes of the meal (standard and large) were used with the large size being 50% bigger than the standard. The food selection during experiments represented a wide range of food properties such as crispiness, moistness, softness/hardness, and tackiness that could impact chewing and swallowing. The provided drink was water. To evaluate the impact of a meal-time conversation on patterns of chews and swallows, the subjects were involved in a dialogue with a member of the research team during the second and fourth visits and ate in silence during the first and third visits. Overall, the human study resulted in 66 h of monitored activity, and captured a total of 10,099 swallows and 46,238 chews. We certify that all applicable institutional and governmental regulations concerning the ethical use of human volunteers were followed during this research. The Institutional Review Board at Clarkson University approved the study, and all subjects signed informed consent forms.

The time series of chews and swallows were used to build prediction models centered on the following hypotheses:

  1. The frequencies of swallowing above a threshold TINGEST or presence of chewing correspond to food ingestion;
  2. The frequencies of swallowing above a threshold TLIQUID > TINGEST- and absence of chewing correspond to liquid ingestion;
  3. The number of chews and swallows is proportional to the ingested mass.

Timescale selection

The predictive models assigned a class label to a time window of certain length (a decision epoch). Every epoch in a subject’s visit was assigned a label from the set {“no intake,” “intake”} indicating the state of food intake. Epochs labeled as “intake” were further labeled from the set {“solid,” “liquid”} thus differentiating between food types. The choice of epoch duration was an important factor defining the time resolution. Naturally, a shorter epoch should provide better time resolution and, in theory, detect shorter events. Such capability would be very important for monitoring of snacking habits. At the same time, the shortest epoch duration is limited by the minimal detectable changes in the predictor variables.

To estimate the duration of a decision epoch we define the instantaneous swallowing frequency: ISFi =60/(titi−1) (sw/min), where t is time of swallow occurrence in seconds, i =2,…, N and N is the total number of swallows. In other words, this measure provides instantaneous information on the number of swallows per minute. A graph of ISF for a subject’s visit (Figure 1a) provides an intuitive illustration of ISF’s predictive abilities. The duration of a decision epoch should be long enough to detect an increase in the number of swallows associated with food consumption and short enough to provide adequate time resolution for detection of food intake. An average of ISF over duration of an epoch is defined as EISFj. To use EISFj as a predictor variable, the duration of a decision epoch will need to be greater than lower bound on ISF for ingestion (established experimentally as the 1st percentile of ISF during food intake ISFFOODLB), which equals ~2 sw/min. Experimental data (Figure 1b) also show that for periods with no ingestion the median ISF ≈ 2 sw/min, thus leading to a conclusion that decision epoch length of 30 s is the most appropriate one, providing the best trade-off between recognition accuracy and time resolution.

Figure 1
Instantaneous swallowing frequency (ISF) is a key predictor for differentiation between intake/no intake of food and solid/liquid food. (a) ISF graph clearly indicates ingestion during the course of a meal. Resting periods (no intake) are labeled 1, intake ...

Group model for detection of food intake

Food ingestion can be identified by swallowing alone or by a combination of chewing and swallowing. Chewing by itself cannot identify intake of liquids and thus is not considered. Recognition of food intake with swallowing frequency as a predictor can be performed on a model (see Supplementary Table 1 online, model 1) that uses Bayes optimal threshold (20) TINGEST, which defines an optimal decision boundary between the classes of “intake” and “no intake.” Food intake is detected if the swallowing frequency is greater or equal than the threshold. This prediction model is based on population-based probability density estimates (21) and therefore is termed a group model. The decision threshold TINGEST = 4 sw/min was established at the intersection of probability density estimates for “no intake” and “solid food” (Figure 1c). Accuracy of prediction was calculated based on the standard statistical definitions of sensitivity and specificity (22). For food intake detection, the group model was capable to achieve an average accuracy of 82% (Table 1).

Table 1
Sensitivity, specificity, and confidence intervals of predictor models

Floating average models for detection of food intake

Accuracy of detection of food ingestion was further improved by taking into account interindividual variability in swallowing rates. The ratio of EISF during resting to EISF during solid food intake (Figure 1d) has a well-defined upper bound of ≈0.5, which means that regardless of the absolute value of resting EISF, the swallowing frequency during food consumption is approximately twice as high. A more accurate floating average model (see Supplementary Table 1 online, model 2) can be formulated with the decision threshold TFLINGEST being a product of a floating average of the swallowing frequency over several epochs and a scaling factor α, which sets the threshold at a certain level above the floating average. This model self-adjusts to individual variations in swallowing rates and thus is expected to perform better than the group model, which is based on population statistics. Training and validation of the floating average model demonstrated 87% accuracy of food intake detection and a reduced variation in prediction rates among subjects.

A floating model can also incorporate chewing as a binary indicator of food intake (see Supplementary Table 1 online, model 3). Although in practical terms this requires an additional sensor that reports on the state of chewing (“chewing” or “no chewing”) the advantage is a higher accuracy of detecting intake of solids. Incorporating chewing increased the average accuracy to 95.5% (Table 1). Such high accuracy for 30-s epochs should be sufficient for detection of most short-term events associated with snacking. An interesting observation is that, with the inclusion of chewing, the decision of food intake for solids relies heavily on the presence of chewing while a higher threshold (α ≈ 1.7) for swallowing frequency captures the consumption of liquids.

Models for differentiation of solids and liquids

The models for differentiation of liquid and solid foods were built using the same principles as models for detection of food intake. The group model (see Supplementary Table 2 online, model 1) based on the optimal Bayes threshold TLIQUID = 12 sw/min (Figure 1c) resulted in 82.2% average accuracy. The floating average model (see Supplementary Table 2 online, model 2) did not significantly improve the average recognition accuracy resulting in 82.5% average accuracy. However, a floating average model that included chewing (see Supplementary Table 2 online, model 3) made a dramatic improvement in accuracy. This could be expected because liquid consumption results in a high swallowing rate easily exceeding the ingestion threshold but it is quite uncommon to chew liquids. The model incorporating chewing produced 93.3% average accuracy. Similar to the model for detecting food ingestion, introducing chews into the decision rule increased the swallowing threshold (α ≈ 2.3) attributing solid food intake to the presence of chewing and intake of liquids to high swallowing rates.

Mass prediction models

The third study hypothesis was tested by showing that chews and swallows can predict the total ingested food mass with reasonable accuracy. The mass prediction model for solid foods is based on a simple linear model using the total number of chews and swallows associated with a period of ingestion as predictors. Predicted mass of solid food was computed as MS=0.5(M¯SWS×NSW+M¯CHEW×NCHEW) where M¯SWS is subject’s average mass per swallow of solid food, NSW is total number of swallows for a period of solid food intake, [M with macron]CHEW is average mass per chew and NCHEW is total number of chews. The M¯SWS and [M with macron]CHEW are individual statistical estimates of the average mass consumed in a single swallow or a chew (see Supplementary Appendix 1 online).

The model for predicting ingested mass of liquids used only swallows as liquids do not involve chewing: ML=M¯SWL×NSW where ML is predicted mass of liquid and NSW is the total number of swallows for liquid intake. However, absence of chewing was not the only reason to have a separate model for liquid intake. Our results show that the average size of a swallow (Figure 2a,b) is significantly larger (P < 0.001) for liquids ( M¯SWL17.8g) than for solids ( M¯SWS6.08g). Therefore, a prediction model taking into account differentiation between solids and liquids is needed for higher accuracy. The population means of mass per swallow and mass per chew are not representative of subject’s individual means due to observed high intersubject variability (Figure 2a–c). This suggests a need for individual- rather than population-calibrated models. Another interesting observation is that mass per swallow and mass per chew have no obvious correlation with subject’s BMI.

Figure 2
Average mass per swallow for solids and liquids and average mass per chew for solids are the key predictors in estimating the mass of ingested food. (a) Average mass per solid food swallow per subject arranged by subject’s BMI. (b) Average mass ...

The accuracy of mass prediction for solid and liquid foods was estimated by sixfold cross-validation (23) on the data from 16 subjects with available four visits. The error was estimated as an absolute value of the percentage difference to the real weight of the food. The mass model of solid food intake achieved 91.79% average accuracy (Figure 3a,b). The mass model for liquid intake achieved 83.76% average accuracy. Lower accuracy of predicting mass of liquids could be attributed to the fact that the liquid ingestion model is based only on swallowing counts while the accuracy of the solid ingestion model is improved by taking chews into account. It should be noted that other parameters of ingestion such as duration of a swallow, total ingestion time, etc. could potentially be factored in to improve accuracy.

Figure 3
The error of mass prediction. (a) Average accuracy of absolute prediction for mass of solid food per subject. Each data point represents an average across sixfolds of training/validation. (b) Average accuracy of absolute prediction for mass of ingested ...

DISCUSSION

The results suggest that accurate capturing of chewing and swallowing events could lead to development of a novel method of accurately measuring human eating behavior and food intake. This in turn could substantially improve our ability to study the eating behaviors associated with obesity.

Based on the reported results, we estimate that measuring chews and swallows can achieve >95% accuracy in detection of food ingestion on short 30-s decision epochs, which should be sufficient to capture short events of snacking; differentiate between solid foods and liquid foods with >91% accuracy; and estimate mass of solid food intake with >91% accuracy and mass of liquids with >83% accuracy. These numbers are likely to improve with development of more sophisticated and self-adjusting individualized prediction models. However, the current accuracy of detecting ingestion is sufficient for reliable behavioral monitoring and providing information on patterns of food intake. Further research is needed to improve accuracy of mass prediction, potentially achieved through incorporation of more features such as time of chewing and duration of swallowing. The advantages of using chews and swallows for monitoring of ingestive behavior include: (i) a more objective means of monitoring of food intake than methods currently available; (ii) the ability to provide an accurate assessment of the daily pattern of eating and drinking with a high time resolution (e.g., 30 s in this study); (iii) information about solid vs. liquid consumption; (iv) accurate estimates of amount of food eaten; and (v) potential for utilizing this method in a wearable noninvasive device.

We also show that detection of food intake and differentiation of liquid/solid food can be performed by monitoring of swallowing alone though the overall accuracy is reduced. This may be important from the perspective of developing a simple biofeedback device to monitor food ingestion and provide feedback in a manner similar to that of a pedometer. Utilizing just one sensor may significantly simplify application of the device and reduce costs. We believe that such a device may be valuable for continuous monitoring of ingestive habits and adaptation of the lifestyle.

The further improvements of this method will likely result in greater accuracy and utility. For example, the microstructure of a meal (i.e., the specific sequence of chews and swallows) may be indicative of the type of food being consumed (23). Alternatively, food type may be reported by a high-tech electronic nose, through a food diary or automatic food photography.

Additional increases in prediction accuracy in detecting chewing, swallowing, and ingested mass may potentially be achieved by involving more of the features characterizing ingestion: time of chewing and duration of swallowing, chewing frequency, etc. to name a few. The algorithms presented here may be combined in a number of ways, which still need to be studied.

In this study, the scores of chews and swallows were obtained manually from the signals captured by video, sound, and other sensors (19). To be practically applicable in free-living conditions, the method of capturing chews and swallows would have to rely on a reduced set of measurements that can be automatically processed by a wearable computing platform (24). An important part in our research is dedicated to building noninvasive wearable sensors and related signal-processing and pattern recognition methods for automatic detection of swallowing instances and periods of chewing, thus not requiring a person to manually score the data. In our publications (19,25), we have shown that swallowing events can be identified by a miniature microphone detecting a specific sound of deglutition and chewing events can be detected by a strain sensor positioned behind the outer ear. These sensor modalities represent just a few of the choices that can be explored for design of an inconspicuous wearable device that would detect and characterize food ingestion by measuring chews and swallows. Such a wearable device can be used for objective monitoring of ingestive behavior of a free-living population. As such a technique would not require a conscious effort on the part of the user, we may expect a reduced observation effect, though this issue will require a separate study.

In summary, we have presented a rationale and preliminary results for a novel method of accurately assessing ingestive behavior in free-living humans based on chewing and swallowing information. This method has the potential to improve our understanding of human eating behaviors associated with obesity.

Supplementary Material

Supplementary files

Acknowledgments

We gratefully acknowledge the support provided by National Institutes of Health grant 5R21HL083052-02 02 and anonymous reviewers for their suggestions on improving the paper.

Footnotes

SUPPLEMENTARY MATERIAL

Supplementary material is linked to the online version of the paper at http://www.nature.com/oby

DISCLOSURE

The authors declared no conflict of interest.

References

1. World Health Organization, Infobase page for Overweight & Obesity (BMI) [Accessed May 2008]. < http://www.who.int/infobase/report.aspx?rid=118&iso=USA&Def_Code=cd.0701&Survey_Year_End=2005&genGraphButton=Generate+Graph>.
2. Carmelli D, Zhang H, Swan GE. Obesity and 33-year follow-up for coronary heart disease and cancer mortality. Epidemiology. 1997;8:378–383. [PubMed]
3. James WP. What are the health risks? The medical consequences of obesity and its health risks. Exp Clin Endocrinol Diabetes. 1998;106 (Suppl 2):1–6. [PubMed]
4. Eckel RH, Krauss RM. American Heart Association call to action: obesity as a major risk factor for coronary heart disease. AHA Nutrition Committee. Circulation. 1998;97:2099–2100. [PubMed]
5. Olshansky SJ, Passaro DJ, Hershow RC, et al. A potential decline in life expectancy in the United States in the 21st century. N Engl J Med. 2005;352:1138–1145. [PubMed]
6. Wisse BE, Kim F, Schwartz MW. Physiology. An integrative view of obesity. Science. 2007;318:928–929. [PubMed]
7. Herbert A, Gerry NP, McQueen MB, et al. A common genetic variant is associated with adult and childhood obesity. Science. 2006;312:279–283. [PubMed]
8. Liu YJ, Liu XG, Wang L, et al. Genome-wide association scans identified CTNNBL1 as a novel gene for obesity. Hum Mol Genet. 2008;17:1803–1813. [PMC free article] [PubMed]
9. Frayling TM, Timpson NJ, Weedon MN, et al. A common variant in the FTO gene is associated with body mass index and predisposes to childhood and adult obesity. Science. 2007;316:889–894. [PMC free article] [PubMed]
10. Hill JO, Wyatt HR, Reed GW, Peters JC. Obesity and the environment: where do we go from here? Science. 2003;299:853–855. [PubMed]
11. Diliberti N, Bordi PL, Conklin MT, Roe LS, Rolls BJ. Increased portion size leads to increased energy intake in a restaurant meal. Obes Res. 2004;12:562–568. [PubMed]
12. Bray GA. How do we get fat? An epidemiologic and metabolic approach. Clin Dermatol. 2004;22:281–288. [PubMed]
13. Hill JO, Peters JC. Environmental contributions to the obesity epidemic. Science. 1998;280:1371–1374. [PubMed]
14. Levine JA, Schleusner SJ, Jensen MD. Energy expenditure of nonexercise activity. Am J Clin Nutr. 2000;72:1451–1454. [PubMed]
15. Chen KY, Acra SA, Majchrzak K, et al. Predicting energy expenditure of physical activity using hip- and wrist-worn accelerometers. Diabetes Technol Ther. 2003;5:1023–1033. [PMC free article] [PubMed]
16. Ainslie PN, Reilly T, Westerterp KR. Estimating human energy expenditure: a review of techniques with particular reference to doubly labelled water. Sports Med. 2003;33:683–698. [PubMed]
17. Fehily AM. Methods of food intake measurement: their uses and abuses. Nutr Bull. 1993;18:25–33.
18. Livingstone MBE, Black AE. Markers of the validity of reported energy intake. J Nutr. 2003;133:895–920. [PubMed]
19. Sazonov E, Schuckers S, Lopez-Meyer P, et al. Non-invasive monitoring of chewing and swallowing for objective quantification of ingestive behavior. Physiol Meas. 2008;29:525–541. [PMC free article] [PubMed]
20. Bowman AW, Azzalini A. Applied Smoothing Techniques for Data Analysis. Oxford University Press; UK: 1997.
21. Shiavi R. Introduction to Applied Statistical Signal Analysis. Academic Press; UK: 2006.
22. Olson DL, Delen D. Advanced Data Mining Techniques. Springer; USA: 2008.
23. Stellar E, Shrager EE. Chews and swallows and the microstructure of eating. Am J Clin Nutr. 1985;42:973–982. [PubMed]
24. Amft O, Tröster G. Recognition of dietary activity events using on-body sensors. Artif Intell Med. 2008;42:121–136. [PubMed]
25. Makeyev O, Sazonov E, Schuckers S, et al. Limited receptive area neural classifier for recognition of swallowing sounds using continuous wavelet transform. Proceedings of the 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBC 2007; Lyon, France. 23–26 August 2007; pp. 3128–3131. [PubMed]