Body fat percentage predicted how badly four smartwatches misread calorie burn

A wrist device senses two things. An accelerometer picks up movement, and an optical sensor infers heart rate from the light the blood in your wrist vessels absorbs and reflects as it pulses ^[5]. Calories burned is the number the hardware cannot sense at all, so the watch predicts it from those signals plus age, sex, height and weight ^[6]. The author, an exercise scientist who worked with colleagues at Florida International University's Medical Photonics Laboratory, writes that the prediction is sensitive to differences in body composition, fitness, movement efficiency and physiological response to exercise ^[3][7].

The protocol was built to make that prediction easy. Recumbent cycling let the team hold intensity steady while reducing wrist movement, which interferes with optical heart rate ^[11]. Each person wore two watches per wrist with placements randomized, and the team found no evidence that one position helped ^[12]. The reference was a metabolic analyzer measuring inhaled and exhaled gases through a mask, which exercise science treats as the most reliable measurement of calorie expenditure ^[10].

With 58 participants each wearing all four devices, the study produces 232 device-sessions, 58 per model ^[17]. Fitbit's 58 include the strangest numbers. After data-quality exclusions its average error was neither systematically high nor low, but seven readings exceeded 450% of the metabolic measurement ^[15]. A reading at 450% of the criterion overstates the burn by a factor of 4.5 ^[18]. If each participant produced one estimate per device, those seven are roughly one Fitbit session in eight ^[19].

The overall inaccuracy was expected; prior research has documented it, and the author reports the strength of the body fat relationship as the finding that surprised the team ^[16]. A flat 20% overestimate can be subtracted, or ignored in favor of tracking the day-to-day direction. An error that grows with body fat percentage cannot be handled that way, because a person whose body fat falls during a training program gets an error that changes size as they go ^[4]. The exposed population is large: about 100 million US adults, nearly 4 in 10, own a smartwatch, and market research puts worldwide wearers near 560 million ^[1][2].

The Conversation account gives the direction of the body fat relationship but not its size, and no result for skin tone, though the sample was recruited to span skin tones as well as body fat percentages ^[8]. So the gradient is a direction. Nobody can yet adjust a watch's calorie figure for a given body fat percentage. The test was one recumbent cycling session, and walking, running and lifting all carry the wrist motion this design deliberately suppressed ^[11]. The sample was 58 Hispanic adults in one city ^[8].