EE Prediction Models
Four out of five published EE prediction models based on ActiGraph devices were assessed, including three models based on proprietary Vector Magnitude Count (VMC) and one based on raw acceleration signals [9, 20–22]. Garcia et al.’s model was not included due to its computational complexity in generating the features [23]. The included models have varied outputs (i.e., physical activity EE, total EE, and oxygen consumption). All outputs were converted to total energy expenditure (TEE) for evaluation against the DLW measurement. Details of the conversion and the original model equations can be found in Table 1.
Table 1
Equations and TEE Conversion of Existing Models
Model (Author) | Equation (units) | TEE Conversion (kcal) |
Model 1 (Nightingale, 2014) [20] | PAEE = (0.000929 × VMC) – 0.284818 (kJ∙min− 1) | TEE (kal∙min− 1) = (PAEE × 0.239* + RMR)/0.9† |
Model 2 (Nightingale, 2015) [21] | PAEE = (0.000245 × VMC) + 0.132379 (kal∙min− 1) | TEE (kal∙min− 1) = (PAEE + RMR)/0.9† |
Model 3 (Tsang, 2018) [9] | TEE = -0.006197602220975 + 0.000000088463104 × BMR2 × VM + 0.000823693371782 × BMR1 + 0.000577607827818 × X (kal∙min− 1) | N/A |
Model 4 (Learmonth, 2016) [22] | Right wrist: VO2 = 0.0022 × VMC + 3.13 (ml/min− 1kg− 1) Left wrist: VO2 = 0.0021 × VMC + 3.14 (ml/min− 1kg− 1) | TEE (kal∙min− 1) = (VO2 /1000‡) × Weight (kg) × 5§ |
Note. In model 3, BMR1 refers to the basal metabolic rate by World Health Organization [24]. BMR2 refers to the basal metabolic rate by Mifflin et al. (1990) [27]. |
PAEE – Physical Activity Energy Expenditure |
RMR - Resting Metabolic Rate |
*Factor used to convert kJ to kcal |
†Factor used to adjust for the diet-induced thermogenesis, which is commonly estimated as 10% of the TEE[18, 25, 26] |
‡Factor used to convert ml to L |
§Calorie equivalent of VO2, 5 kcal/L |
In addition, a random forest (RF) model using raw acceleration signals from ActiGraph activity monitors was developed by aggregating data from two published studies that shared similar inclusion/exclusion criteria and study protocols [27]. In both studies, participants were fitted with a K4b2 portable metabolic cart (COSMED USA Inc., Concord CA, USA) for criterion TEE measurement, and an ActiGraph activity monitor on the dominant wrist and performed a series of activities of daily living and exercises. The PA measured by the activity monitor and TEE measured by the metabolic cart were extracted and cleaned. A set of features, including features based on raw acceleration signals and demographic variables, was generated and then selected using Pearson’s r. We used the “grid search” approach to tune the hyperparameters for the RF model. The Leave-One-Out Cross-Validation was used to evaluate the model performance of each hyperparameter combination. The hyperparameter combination that achieved the best mean squared error (MSE) was used to train the final RF model using all participants’ data for the subsequent free-living field validation. More details about the model development can be found in Supplementary Appendix A.
Free-Living Validation via DLW
This study was conducted at the University of Pittsburgh Endocrinology and Metabolism Research Center at Montefiore Hospital, Pittsburgh. It was approved by the Institutional Review Boards of the VA Pittsburgh Healthcare System and the University of Pittsburgh. We included participants who were: 1) between 18 and 65 years, 2) having an SCI, 3) at least one-year post-injury, 4) using a manual wheelchair as a primary means of mobility (≥ 80% of weekly mobility), 5) medically stable, 6) living within 1 hour of the research site, 7) being able to tolerate sitting for three hours, and 8) being able to self-catheterize for the doubly labeled water activities. We excluded participants who were: 1) having any active pelvic or thigh wounds, 2) having a history of cardiovascular disease, 3) having high blood pressure, seizures, lung disease, fainting/dizziness, shortness of breath at rest or with usual activities, or unusual fatigue, and 4) pregnant. Convenience sampling was used for recruitment.
As the purpose of this paper was to validate EE prediction models, only EE-related instruments and data were reported.
Study Protocol
The study consisted of two lab visits with a one-week home trial in between. Participants’ criterion TDEE and upper extremity kinematic data during the study period were measured by the DLW and the ActiGraph activity monitor (GT9X Link), respectively. The DLW dose was approximately 3.5 oz (actual weight of the DLW was noted for each participant). After drinking the dose, the bottle was rinsed with 3 oz tap water, which the subject then drank. The flowchart (Fig. 1) shows the timeline. The anthropometric measurements of height were conducted in a supine position with a tape measure to the closest 1 cm and converted to inches. The participant’s weight was calculated by subtracting the wheelchair weight from their total weight measured by a wheelchair scale (Detecto,Webb City MO, USA) and rounded to the nearest 0.1 lb. Body composition was measured via the dual-energy X-ray absorptiometry scan (DEXA) using the GE Lunar iDXA (GE Healthcare, Wauwatosa WI, USA). Their resting metabolic rate (RMR) was measured in the supine position on a bed for 30 minutes in a dimly lit, quiet room using the TrueOne stationary metabolic cart (Parvo Medics, Salt Lake City UT, USA) following 12 hours fast. They were instructed not to talk and not to fall asleep during the measurement. More details about the study instruments can be found in Supplementary Appendix B.
All three urine samples from each participant were refrigerated and kept at the lab until the whole study concluded. They were then sent to the AdventHealth Orlando, Translational Research Institute (Orlando, FL.) for processing and analysis.
[Figure 1]
Data Collection and Preparation
The raw acceleration outputs and the minute-by-minute VMC were downloaded and extracted using ActiGraph’s proprietary software ActiLife (v6.11.9) from the GT9X Link. After data cleaning, the ActiGraph data was segmented into valid wear, non-wear, and sleep periods based on the self-report device wear and activity log, and the wear time validation function in the ActiLife software. The developed RF model, as well as the four existing models were then applied to the valid wear period, while lab-measured RMR was used to compensate for non-wear and sleep periods for estimating the TDEE. Data collection and preparation procedure flowchart is presented in Fig. 2, while more details can be found in Supplementary Appendix C.
[Figure 2]
Data Analysis
RF model performance analysis
The performance of the final model was presented by the mean and SD of the mean absolute error (MAE) and MAPE across all participants. The result was segregated into three activity intensity levels based on the metabolic equivalent task (MET): 1) Sedentary (MET < 1.5), 2) Light (1.5 ≤ MET ≤ 3), 3) Moderate to Vigorous Physical Activity (MVPA) (MET > 3). The importance of each feature was estimated by averaging their importance across all decision trees in the model [28]. The Bland-Altman (BA) plot was used to further evaluate agreements and bias between estimated and criterion EE.
Field validation
Any participant’s data with non-wear time larger than 20% of the trial length was excluded. Due to the small sample size and preliminary nature of this study, descriptive analysis was provided to evaluate the field validity of the models. The final estimated TDEE was validated against the criterion TDEE measured by the DLW. The overall MAE and MAPE were assessed. All data analyses in this paper were performed by SPSS version 26 (IBM, Armonk NY, USA) and Microsoft Excel (Micosoft, Redmond WA, USA).