TL;DR
This paper introduces a novel wearable system that accurately estimates metabolic energy expenditure in real-time using inertial measurement units (IMUs) on the shank and thigh. This system significantly outperforms existing wearables, including smartwatches, by leveraging lower-limb kinematics, paving the way for more effective tools for weight management and large-scale activity monitoring.
What Problem Does This Paper Address?
Physical inactivity is a major global health concern, contributing to significant mortality. To combat this, objective and scalable methods for measuring physical activity are needed. While gold-standard methods like respirometry and doubly labeled water exist, they are impractical for everyday use. Current portable solutions, such as smartwatches, suffer from substantial errors, particularly in estimating time-varying activities which constitute a significant portion of daily movement. The challenge lies in developing a wearable system that is accurate, portable, low-cost, and can reliably estimate energy expenditure for diverse individuals and activities without requiring subject-specific calibration.
Methods
The core of this work is a "Wearable System" that utilizes a data-driven approach to estimate energy expenditure. Key methodological aspects include:
- Sensor Selection: Through rigorous experimentation, the researchers identified that two IMUs, one on the shank and one on the thigh of a single leg, provided the most informative data for energy expenditure estimation. This selection process prioritized sensors that were low-cost, unobtrusive, and allowed for natural movement.
- Data-Driven Model: A linear regression model was employed, trained on data from various activities. Crucially, the model segments data by stride, allowing for time-invariant inputs that are computationally efficient.
- Lower-Limb Kinematics: The system's novelty lies in its focus on lower-limb movement. The authors argue that leg kinematics are more indicative of energy expenditure than wrist or trunk movements, and physiological signals (like heart rate) which have a delayed response.
- Robustness to Orientation: To improve the model's robustness, synthetic data with random rotations of the IMUs was used during training, addressing potential errors due to sensor placement variations.
- Real-time Estimation: The system is designed for real-time computation, with estimates generated in approximately 0.01 seconds per stride.
Datasets
The study utilized several datasets for training and validation:
- Training Data: This dataset was compiled from previous experiments and new data collected for stair climbing and biking. It included data from 13 young, healthy subjects performing various steady-state conditions of walking, running, stair climbing, and biking.
- Validation Data: A new dataset was collected from 24 diverse subjects (15 men, 9 women) who had not participated in previous experiments. These subjects performed new steady-state conditions at intermediate speeds and four new time-varying conditions for walking, running, stair climbing, and biking. This diverse cohort and novel conditions aimed to test the system's generalizability.
Key Insights and Ideas
- Lower-limb is King: The most significant insight is that focusing on lower-limb kinematics, specifically from the shank and thigh, provides substantially more accurate energy expenditure estimates than wrist or trunk-based sensors, or even physiological signals like heart rate. This is likely because the legs contribute the most to overall energy expenditure during locomotion.
- Stride Segmentation is Crucial: Segmenting sensor data by stride, rather than relying on continuous acceleration counts or physiological signals, allows the model to capture the cyclical nature of movement and leads to more accurate and timely estimations, especially during time-varying activities.
- Careful Sensor Selection Trumps Sensor Fusion: Counterintuitively, the careful selection of just two IMUs on the leg outperformed models that used a broader range of sensors. This highlights the importance of what data is collected, not just how much.
- No Subject-Specific Calibration Needed: The system's ability to achieve low error without requiring individual calibration is a major step towards practical, large-scale deployment.
Key Results
- Cumulative Error: The Wearable System achieved a cumulative energy expenditure error of 13% across common activities (steady-state and time-varying) when evaluated with new subjects and conditions.
- Comparison to Smartwatches: This is significantly lower than the 42% error for a standard smartwatch and 44% for an activity-specific smartwatch.
- Time-Varying Activity: The system demonstrated a 23% error over time for time-varying activities, compared to 46-105% for other methods.
- Steady-State Accuracy: For steady-state conditions, the Wearable System achieved a 13% steady-state error, about half that of the next best wearable method.
- Usability: Subject surveys indicated high usability and comfort for the Wearable System, scoring in the 90th percentile for System Usability Scale.
Limitations and Future Work
- Training Data Dependency: While the system doesn't require subject-specific calibration for testing, the model itself is trained on specific activities. To estimate energy expenditure for entirely new activities, additional training data would be needed.
- Aerobic Threshold: The study focused on activities within the aerobic threshold, as accurately measuring energy expenditure beyond this point is challenging with current ground truth methods.
- Hardware Prototype: The current Wearable System is a research prototype with a bulky microcontroller and wired sensors. Future work will focus on refining the hardware for practical, large-scale deployment, potentially by integrating sensors into smartphones or clothing.
- Diverse Activities: While the system performs well on common activities, extending its capabilities to activities with significant upper-limb involvement or those beyond the aerobic threshold remains an area for future research.
- Personalized Basal Metabolic Rate: For quiet standing estimation, the system uses a scaled basal metabolic rate. Incorporating more personalized basal metabolic rate estimations could further improve accuracy.
Summary updated