Predicting resting energy expenditure: a critical appraisal

The most commonly used prediction models for resting energy expenditure (REE) are Harris-Benedict (1919), Schofield (1985), Owen (1986), and Mifflin-St Jeor (1990), based on height, weight, age and gender, and Cunningham (1991), based on body composition. Here, the five models are compared with reference data, consisting of individual REE measurements (n = 353) from 14 studies, covering a large range of participant characteristics. For white adults, prediction of REE with the Harris-Benedict model approached measured REE most closely, with estimates within 10% for more than 70% of the reference population. Sources of differences between measured and predicted REE include measurement validity and measurement conditions. Importantly, a 12- to 14-h overnight fast may not be sufficient to reach post-absorptive conditions and may explain differences between predicted REE and measured REE. In both cases complete fasting REE may not have been achieved, especially in participants with high energy intake. In white adults, measured resting energy expenditure was closest to predicted values with the classic Harris-Benedict model. Suggestions for improving resting energy expenditure measurements, as well as prediction models, include the definition of post-absorptive conditions, representing complete fasting conditions with respiratory exchange ratio as indicator.


INTRODUCTION
Resting energy expenditure (REE) reflects the maintenance component of daily energy expenditure (DEE). Therefore, measurement of REE must meet four conditions: (1) the participant is awake; (2) in a thermoneutral environment to avoid heat production for maintenance of body temperature; (3) in the post-absorptive state to avoid diet-induced thermogenesis; and (4) in rest to avoid activity-induced thermogenesis [1]. Resting energy expenditure is often used in order to evaluate physical activity and reported energy intake [2]. The physical activity level (PAL) of an adult can be calculated by expressing DEE as a multiple of REE: PAL = DEE/REE [2]. Typical PAL values are: 1.40-1.69 for a sedentary or light-active lifestyle; 1.70-1.99 for active or moderately active lifestyles; and 2.00-2.40 for a vigorously active lifestyle [2]. Reported food intake can be evaluated accordingly, by expressing calculated energy intake as a multiple of measured or predicted REE [3]. REE is primarily determined by body composition. Models for REE-prediction are based on measured body composition or on body characteristics determining body composition, generally including height, weight, age and gender. Thus, models for energy requirement are based on predicted REE and estimated PAL [4].
Here, REE-prediction models as based on estimated body composition and on measured body composition, are compared with published data on measured REE as a reference. The reference REE data (n = 353) were obtained from 14 studies that nearly all used the Omnical (Maastricht Instruments, Maastricht, The Netherlands) for their measurements. In a recent study testing the validity of four commercially available metabolic carts for assessing REE (5), the Omnical had appeared to be the only currently available and acceptable instrument with a measurement error lower than 2% [5]. Special emphasis in the discussion is on sources of differences between measured and predicted REE, including measurement conditions and measurement validity.

Reference resting energy expenditure data
Reference data to compare predicted REE with measured REE include data from published doubly labeled water studies with participation of Maastricht University. These include individual data on height, weight, age, and gender, and fat-free mass (FFM) and fat mass (FM) based on isotope dilution, REE, and DEE [6][7][8][9][10][11][12][13][14][15][16][17][18][19]. All REE measurements were performed with a ventilated hood, in the post-absorptive state, in the morning after waking up and before breakfast, at least 12 h after the last food intake the night before. Data were selected from adult participants (age >18 y), under neutral energy balance and before any study intervention. The data did not include participants with muscle wasting or diseases affecting REE.

Models based on estimated body composition
The most common REE-prediction models based on estimated body composition are, in historical sequence: Harris-Benedict [20], Schofield [21], Owen [22,23], and Mifflin-St Jeor [24]. Here, the Harris-Benedict model [20] and the Owen model [22,23] were derived by combining the original data for women and men (Table 1). Schofield [21] uses separate models for women and men and for participants aged 18-30, 30-60 and over 60 y.

Model based on measured body composition
The most common REE-prediction model based on measured body composition is the Cunningham model [25]. It was calculated from the intercept and slope weighted by sample sizes, of seven studies with 104 to 483 participants each [25]. The total number of participants, was 1483 women and men, age≥18 y, including adults of different weight for height (Table 1).

Reference resting energy expenditure data
The reference group included 353 participants, 154 women and 199 men ( Table 2). Participant age ranged between 19 y and 80 y, similar for women and men, with similar means and standard deviations as well. Mean values for height, body mass, body composition, REE and DEE, showed differences between women and men as expected. The reference group included 72 participants with obesity (BMI ≥ 30 kg/m 2 ), 46 women and 26 men. Mean PAL values were identical to a study on physical activity levels in people from affluent societies [26].

Models based on estimated body composition
Similarity between predicted and measured REE, was highest for the Harris-Benedict and Schofield models as shown by Bland-Altman plots (Fig. 1), and by the frequency distribution of the number or percentage of participants with predicted REE within ±10% of measured REE (Fig. 2, Table 3). Similarity between predicted and measured REE was comparably high for subgroups of only women or men, and for participants with BMI < 30 kg/m 2 or BMI ≥ 30 kg/m 2 . Similarity between predicted and measured REE for the Schofield model was lower for participants with BMI ≥ 30 kg/m 2 . Similarity between predicted and measured REE for the Owen and Mifflin-St Jeor models, based on populations with a relatively high number of participants with BMI ≥ 30 kg/m 2 , was highest for participants with BMI ≥ 30 kg/m 2 as well.

Model based on measured body composition
Similarity between predicted and measured REE for the model based on measured body composition was not higher than for the models based on estimated body composition (Fig. 2, Table 3). Despite inclusion of participants with BMI ≥ 30 kg/m 2 in the model based on measured body composition, the model showed an estimation similarity of only 50% compared to the reference population participants with BMI ≥ 30 kg/m 2 .

DISCUSSION
For white adults, prediction of REE with models based on estimated body composition like the Harris-Benedict and Schofield models approached measured REE most closely, with estimates within 10% for more than 70% of the reference population. For some ethnicities like South Asian Indians or participants with an exceptional high physical activity level like athletes, prediction of REE with models based on measured body composition more closely approach measured REE [27,28]. The physiological relevance of numerical values of the coefficients in the models based on estimated body composition is difficult to interpret [29]. In general, coefficients for height and weight, as indicators for body size are positive. The coefficient for age is negative, reflecting the decrease in FFM and the increase in FM with increasing age. The higher FFM in men compared to women of the same body mass, height and age is reflected in the positive sex coefficient for men, being zero for women (Table 1).
A literature comparison of the validity of prediction models for REE with measured REE, showed that the Mifflin-St Jeor model resulted in predicting REE within ± 10% of measured REE in more participants with BMI ≥ 30 kg/m 2 than any other model [30]. The same four prediction models, identified as the most commonly used in clinical practice, were included in the comparison. A comparison of more than 20 models, including the four models of the current evaluation, in US and Dutch participants with BMI 25-40 kg/m 2 , concluded that REE estimated by the Mifflin-St Jeor model was closest to measured REE [31]. However, in the current analysis the classic Harris-Benedict model, based on a population with only six participants with BMI ≥30 kg/m 2 , performed equally well for all participants, and for the subgroups as defined.
The REE-prediction models in the current analysis, as based on estimated body composition of white participants, do not apply for participants with another body build. As an example, several studies showed that the models overestimate REE relative to measured REE for Asian participants [27,32]. REE-prediction models based on measured body composition do better for application in population groups worldwide [33].
The Cunningham model, based on measured body composition, was calculated from five studies that included participants with BMI < 30 kg/m 2 and with BMI ≥ 30 kg/m 2 . One study included only participants with BMI ≥ 30 kg/m 2 . Two of the studies revealed FM as an additional significant body composition factor in the regression analysis. The coefficient for FM was about one fifth of the coefficient for FFM, and it was concluded that a unique contribution of FM to the regulation of REE requires further elucidation [25]. Later studies confirmed the impact of FM on REE, without providing an explanation [34,35].
Predicted REE with models based on measured body composition is closer to measured REE than predicted REE with models based on estimated body composition for most population groups Difference between measured and predicted value (%) Cunningham Fig. 2 Frequency distribution of the % difference between predicted and measured resting energy expenditure for 353 participants. Five prediction models are used: Harris-Benedict [20]; Schofield [21]; Owen [22,23]; Mifflin-St Jeor [24]; and Cunningham [25]. Black bars indicate predicted values within ± 10% of measured values. [33]. However, the Cunningham model showed a significant overestimation of REE relative to measured REE in master athletes, characterized by a relatively high FFM [28]. One explanation is a negligible contribution of muscle mass to REE under resting conditions. In a 40-week exercise training study, inducing a 2.8 ± 1.6 kg increase in FFM (p < 0.001), REE did not increase [36].
Measurement conditions possibly have not met the four defined conditions, during the measurements for the REE prediction models as well as during the measurements of the reference population. Out of the four conditions mentioned in the introduction, those of greatest importance are rest to avoid activity-induced thermogenesis and the post-absorptive state to avoid diet-induced thermogenesis [20].
Rest during REE measurement can be assured by an automatic record of all movements, even those imperceptible to a trained observer [20]. Post-absorptive conditions for REE measurement are typically defined as a 12-to 14-h overnight fast [37]. Diet induced thermogenesis shows a diurnal pattern with a maximum after the evening meal and a minimum during overnight sleep [38]. However, REE increase in the morning after waking-up and before food consumption likely includes thermogenesis from consumed nutrients the day before. Substrate oxidation after waking-up and before food consumption, still reflects macronutrient composition from the previous diet [39]. Thus REE, measured after a 12-to 14-h overnight fast, does not reflect REE under complete postabsorptive conditions.
A clear indication for a longer interval than a 12-to 14-h overnight fast to reach post-absorptive conditions is a study on the influence of feeding frequency on nutrient utilization. Participants skipping breakfast showed a decrease in respiratory exchange ratio, indicating the transition to fasting-induced lipolysis, only 15-to 18-h after the last meal the night before [40]. Thus, a 12-to 14-h overnight fast may not be sufficient to reach post-absorptive conditions.
The so far unexplained contribution of FM to the variation of REE, in body composition based models, can theoretically be explained by the measurement of REE under non-postabsorptive conditions as well. Respiratory exchange ratio measured after a 14.5-h overnight fast was positively correlated to BMI, as a proxy of FM, in participants fed to estimated energy balance [41]. The positive correlation was explained by delayed food processing in participants with higher energy requirement and thus a higher energy intake to maintain energy balance [41]. Thus, an increased REE by residual diet induced thermogenesis may explain the positive contribution of FM to variation in of REE in body composition based models as specifically observed at higher grades of adiposity [35].
Standardization of REE measurements under post-absorptive conditions can be improved by inclusion of an additional criterion for fasting conditions as indicated by the respiratory exchange ratio. Participants skipping breakfast showed an increase in respiratory exchange ratio directly after waking-up, to a value close to 0.84, reflecting the macronutrient composition from the previous diet [40]. Subsequently, the respiratory exchange ratio decreased below 0.80, indicating the transition to fasting-induced lipolysis, at 15-to 18-h after the last meal the night before [40]. Thus, it is suggested to define the post-absorptive state for REE measurements by inclusion of the value of the respiratory exchange ratio.
Finally, adaptive thermogenesis as already observed after very modest reductions of energy intake [42], can possibly be attributed to the difference in fasting conditions between the baseline REE and REE during energy restricted weight-loss.
In conclusion, when comparing measured REE with predicted REE for a group of white adults covering a large range of participant characteristics, differences were smallest for the classic Harris-Benedict model. Important factors to improve bias and accuracy are measurement validity and measurement conditions. For the latter further research is required to define post-absorptive conditions. Table 3. Percentage participants with difference between predicted and measured resting energy expenditure within ±10% for five prediction models, with values >70% in bold.