Sensing leg movement enhances wearable monitoring of energy expenditure

Physical inactivity is the fourth leading cause of global mortality, highlighting the critical need for accurate and practical physical activity monitoring tools. Current methods, such as self-reported surveys and step counters, suffer from significant inaccuracies. Laboratory-based techniques like respirometry and doubly labeled water provide accurate measurements but are too expensive and inconvenient for large-scale, everyday use. Smartwatches and activity trackers offer portability but lack the precision needed for reliable energy expenditure estimation, particularly during time-varying activities which constitute a significant portion of daily activity. The existing wearable methods often rely on heart rate, wrist, or trunk kinematics, which may not accurately capture the energy expenditure of lower-limb muscles, the primary contributors to daily energy expenditure. The researchers hypothesized that a data-driven method, focusing on lower-limb kinematics segmented by stride, and devoid of subject-specific training data could significantly improve the accuracy of energy expenditure estimation compared to state-of-the-art methods.

The paper reviews existing methods for monitoring energy expenditure, highlighting their limitations. Self-report surveys are unreliable and lack precision. Pedometers and smartphones provide step counts but make inaccurate assumptions about activity intensity. Laboratory-based methods, while accurate for steady-state activities, are not suitable for everyday use due to cost and invasiveness. Data-driven approaches using wearable sensors offer potential but often rely on subject-specific training data, limiting their generalizability. Existing wearable activity monitors and smartwatches often show substantial errors (30-93%) when evaluated with new subjects, largely attributed to their reliance on heart rate and upper-body kinematics. The delayed response of physiological signals to changes in energy expenditure also contributes to inaccuracies, particularly during time-varying activities. The literature strongly suggests a need for a more accurate, portable, and generalizable method for measuring energy expenditure.

The study involved four main experiments. The first experiment aimed to evaluate various data-driven methods and select the most informative sensors for energy expenditure estimation. Thirteen healthy young adults performed various steady-state activities (walking, running, sideways walking, backward walking, hopping, and loaded walking) while wearing a comprehensive set of wearable sensors (respirometry, heart rate monitor, EMG, IMUs, and force-sensing insoles). A data-driven model using all sensor data segmented by stride showed the lowest error. Sensor selection experiments revealed that using IMUs on the shank and thigh yielded the best results, minimizing error while maximizing practicality. The second experiment expanded to include time-varying conditions, such as periodic transitions between walking and running, to assess the models' ability to capture dynamic changes in energy expenditure. The third experiment involved collecting additional data on stair climbing and biking to train a more robust data-driven model for the Wearable System. This model was specifically designed to be robust to sensor orientation variations through the use of synthetic data augmentation. Finally, the fourth experiment validated the Wearable System with a diverse group of 24 new subjects (15 men, 9 women; age 34.8±11.6 years) performing various steady-state and time-varying activities under new conditions. The Wearable System was compared against several methods: a smartwatch, an activity-specific smartwatch, a heart rate model, respirometry (per-breath and fast-estimated), and interpolated respirometry (for time-varying conditions). Usability and comfort surveys were also conducted on a subset of subjects.

The Wearable System, using shank and thigh IMUs, demonstrated significantly lower errors than state-of-the-art methods across all activities. During steady-state conditions, it achieved a 13% cumulative energy expenditure error, compared to 42% for a standard smartwatch and 44% for an activity-specific smartwatch. The system's ability to capture energy expenditure from the start of steady-state conditions and its accuracy during time-varying activities highlight its advantages. The analysis of the linear regression model used in the Wearable System revealed that gyroscope data and stride information were crucial for accurate estimation. Usability surveys showed a high System Usability Scale (SUS) score (80.9/100) and positive feedback on comfort, indicating the potential for practical application. The study rigorously evaluated the Wearable System's performance across a diverse range of subjects and conditions, demonstrating its robustness and generalizability.

The superior performance of the Wearable System is attributed to its use of lower-limb kinematics, segmented by stride, which are more directly related to the energy expenditure of the leg muscles. The system's low error in both steady-state and time-varying conditions is a significant improvement over existing methods, which often suffer from delayed responses and drift in their estimates. The inclusion of heart rate as an input did not improve accuracy, suggesting that lower-limb kinematics are more informative for energy expenditure estimation. The study's rigorous evaluation with a diverse group of subjects increases the confidence in the Wearable System's generalizability. The findings suggest that carefully selected sensors and data processing techniques, such as stride segmentation, are crucial for improving the accuracy of wearable energy expenditure monitoring.

The Wearable System represents a significant advancement in physical activity monitoring. Its superior accuracy, portability, and low cost make it a promising tool for weight management programs and large-scale epidemiological studies. Future work could focus on miniaturizing the system, integrating it into existing wearable technology (such as smartphones or clothing), and expanding the range of activities monitored. Further research should explore the potential of this approach in diverse populations and clinical settings, potentially paving the way for personalized energy balance systems and improved public health interventions.

The study primarily focused on common activities within the aerobic threshold. Extending the system to encompass a wider range of activities, particularly those with significant upper-body involvement or exceeding the aerobic threshold, would enhance its applicability. The training data for the Wearable System primarily included data from young, healthy adults. While the validation involved a diverse group of subjects, further research with more diverse populations (different ages, fitness levels, etc.) is warranted. Finally, the study used a prototype device, and further development is needed to create a commercially viable product.