One hip-worn sensor estimates walking energy burn using joint mechanics and a tiny neural network
A single sacrum IMU plus a small neural network estimated walking energy expenditure in 13 healthy men (RMSE 0.69 W/kg), though only in a lab treadmill setting.
Summary
Measuring how much energy a person burns while walking normally requires a lab mask system (indirect calorimetry). This engineering study asked whether one inertial sensor on the lower back (sacrum) could replace it. The authors showed that stance-leg sagittal joint power accounts for about 85% of positive whole-body joint power and correlates strongly with it (R ≈ 0.97). A small neural network predicted leg joint angles and torques from the sensor, which were converted to energy expenditure with efficiency weights and a regression scale factor. In an independent test group of 13 healthy young men on a treadmill, error was RMSE 0.69 W/kg. This is a validation of a measurement tool, not a health intervention. It was tested only in healthy adults, mostly men, on a treadmill at set speeds.
Detailed Summary
Energy expenditure during walking is a useful marker for rehabilitation, exercise prescription and daily activity tracking, but the gold standard, indirect calorimetry, needs a face mask and lab equipment. Existing wearable methods often need several sensors or large calorimetry-labeled training sets. This study tests a leaner, physics-guided alternative: one inertial measurement unit (IMU) at the sacrum.
The mechanism is biomechanical rather than biological. Using whole-body motion capture and force plates in 8 healthy adults walking overground at five speeds, the authors found the stance leg produced 84.6% of positive and 76.2% of negative whole-body joint power. Sagittal-plane (fore-aft) power of the stance leg correlated with whole-body joint power at R = 0.97 (0.98 after efficiency weighting, using coefficients of 4.00 for positive and -0.83 for negative work). The swing leg (R = 0.92) and upper body (R = 0.77) tracked less closely.
The team then trained a small neural network (one hidden layer of 20 nodes) on 9 healthy men to predict ankle, knee and hip angles and torques from sacral displacement and treadmill speed. Joint power was derived from those outputs, efficiency-weighted, and scaled by a regression factor to net energy expenditure. Calorimetry was the ground truth at five speeds from 1.0 to 1.75 m/s. On an independent test group of 13 healthy men, error was RMSE 0.69 W/kg. The abstract reports very consistent scaling across speeds and subjects (coefficient of variation under 2%). The authors also compared it with a direct transformer model that maps IMU signals to energy expenditure without a biomechanical intermediate, and say the biomechanically informed approach was more accurate and robust. The full text provided was truncated, so the exact comparison figures are not available here.
The implication is a simple, interpretable wearable method that needs little calorimetry-labeled data. Several caveats matter. Participants were young, healthy and almost entirely male. Walking was on a treadmill at prescribed speeds, and the model used the set treadmill speed as an input rather than estimating it from the sensor. Sagittal power is a proxy and omits frontal-plane work, which was up to about 19% of stance-leg power. Accuracy in older adults, people with gait impairment, slopes, or free-living walking has not been shown. This work says nothing about health outcomes or lifespan.
Key Findings
- Stance leg generated 84.6% of positive and 76.2% of negative whole-body joint power in 8 healthy adults walking overground.
- Stance-leg sagittal power correlated with whole-body joint power at R = 0.97 (0.98 efficiency-weighted); swing leg 0.92, upper body 0.77.
- A one-hidden-layer neural network plus efficiency weights estimated energy expenditure with RMSE 0.69 W/kg in 13 independent healthy men.
- Frontal-plane work added 15-19% of stance-leg power, so the sagittal-only proxy omits part of the mechanical cost.
Methodology
Three healthy-adult datasets: 8 people for overground segmental power analysis, 9 men for training, and 13 men for independent testing on a treadmill at five speeds (1.0-1.75 m/s). Ground truth was net energy expenditure from indirect calorimetry. Joint power came from inverse dynamics, and a small ANN estimated it from sacral IMU data, with leave-one-subject-out validation. It was compared with a direct transformer model.
Study Limitations
Samples were small and homogeneous (healthy, young, nearly all male), with treadmill-only testing at prescribed speeds that were fed to the model. Sagittal-plane power is a proxy that ignores frontal and transverse work, and the net EE ground truth depends on calorimetry assumptions. Performance on slopes, in free-living walking, in older or impaired gait, and in women is unknown.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe:
