39-LB: Estimating Basal Metabolic Rate in People with Diabetes

Abstract

Listen

Translate article

Objective: Health care settings and self-care programs use Basal Metabolic Rate (BMR) to determine caloric need. BMR is difficult to measure directly, so many formulas have been published to estimate BMR, each derived from relatively small samples. We fit two regressions to a large sample of BMR estimates from the Apple Watch for people with diabetes and prediabetes, and compared the results to 10 common BMR formulas. Method: Data included self-reported height, age, sex, weight, diabetes type, and over 500,000 daily values of BMR and caloric expenditures from Apple Watch, from 3,071 people (56% male, 14.5% female; 29.5% unreported; 11.4% type 1, 58.9% type 2, 15.3% prediabetes, 14.4% other diabetes). We regressed to Apple Watch BMR using 5-fold cross-validation, first with height, weight, age and sex only, then adding 30-day calorie expenditure, and measured the hold-out accuracy. Results: The 10 published formulas we evaluated predicted Apple Watch BMR with an average RMSE of 354 ± 67 calories and an average R2 of 0.726 ± 0.074. The Harris-Benedict equation showed the closest correspondence to Apple Watch BMR, with RMSE 287 calories and R2 0.760. Our regression without energy expenditure yielded BMR = -112.2 + 1.08w + 3.89h - 51.6g + 0.13w2 + 0.0467h2 - 0.00041w3 - 0.00017h3, and with energy expenditure, BMR = -232 + 1.66w + 4.01h - 46.9g + 0.13w2 + 0.045h2 - 0.000417w3 - 0.00017h3 + 0.0091e, where w=weight (kg), h=height (cm), g=gender(-1=male, 1=female, 0=unreported) and e = 30-day energy burn (kcal). The accuracies were RMSE 280 calories, R2 0.766; without energy expenditure and RMSE 274 calories, R2 0.780 with energy expenditure. Conclusion: Regressions to more than a half-million samples from over 3,000 people with diabetes predicted Apple Watch BMR estimates with lower error than common published formulas. The most accurate regression included 30-day energy burn, reducing RMSE by 13 calories (4.5%) and increasing R2 by .02 vs. the closest common formula. These results recommend consideration of activity measures when estimating BMR. Disclosure Y. Wexler: None. D. Goldner: Employee; Self; One Drop.