TL;DR
This paper presents data-driven models capable of rapidly estimating human energy expenditure during walking using wearable sensor data (ground reaction forces and muscle activity). These models achieve accuracy comparable to traditional, cumbersome indirect calorimetry, paving the way for more personalized and responsive assistive devices and rehabilitation strategies.
What Problem Does This Paper Address?
Accurately measuring a person's energy expenditure (how much energy their body uses) is crucial for understanding the effort involved in walking, especially for individuals with mobility impairments. This information is vital for designing effective gait retraining programs and optimizing assistive devices like exoskeletons. However, the gold standard for measuring energy expenditure, indirect calorimetry, is expensive, requires specialized equipment, and takes several minutes to collect reliable data for each condition. This makes it impractical for real-time adjustments or evaluating many different conditions quickly. While wearable sensors offer a promising alternative, their accuracy in estimating energy expenditure, particularly for novel conditions or individuals not included in the initial training data, has not been thoroughly evaluated. This paper aims to bridge this gap by developing and validating data-driven models that can estimate energy expenditure in near real-time using readily available sensor data.
Methods
The researchers developed two main types of models:
- Per-Gait Cycle Models: These models estimate energy expenditure for each individual gait cycle (roughly one second of walking). They utilize linear regression and neural networks. The input data (ground reaction forces and electromyography - EMG - signals from leg muscles) were segmented by gait cycle and then further processed into fixed-size bins, with each bin's data averaged.
- Fixed Time Interval Models: For activities without clear periodic features (like a distinct gait cycle), a recurrent neural network (specifically, a Long Short-Term Memory or LSTM variant) was employed. This model processed data over fixed time intervals (four seconds in this study) to capture temporal dependencies.
The models were trained using ground reaction forces (measured by pressure-sensing insoles) and EMG signals (measured by electrodes on the leg muscles). The "ground truth" for energy expenditure was obtained using traditional indirect calorimetry.
Datasets
The study utilized two distinct datasets:
- Assisted Walking Dataset: This dataset involved subjects walking with a unilateral ankle exoskeleton that provided various assistance profiles. This dataset was used to evaluate models for conditions with similar characteristics to those encountered in exoskeleton optimization.
- Inclined Loaded Walking Dataset: This dataset focused on unassisted walking with varying loads (0-30% of body weight) and inclines (0-10% grades). This dataset was used to assess models across a wider range of energy expenditure values.
These datasets were used to evaluate the models across three "use cases":
- Novel Condition: Testing new walking conditions for subjects with some existing training data.
- Novel Subject: Testing all conditions for a completely new subject not included in the training data.
- Both-Novel: Testing new conditions for a new subject, representing a more challenging generalization scenario.
Key Insights and Ideas
- Wearable Sensors as a Viable Alternative: The study demonstrates that data-driven models using ground reaction forces and EMG signals can achieve accuracy comparable to indirect calorimetry, suggesting that wearable sensors can indeed replace traditional methods for many applications.
- Generalization is Achievable: The models showed a promising ability to generalize to new subjects and conditions, particularly when the new conditions were similar to those in the training data. This is a critical step towards real-world applicability.
- Feature Importance Varies: The importance of different input features (ground reaction forces vs. EMG) depended on the walking condition. Ground reaction forces were more critical for inclined and loaded walking, likely capturing more comprehensive information about the task.
- Ordering Matters: Beyond just predicting the exact energy expenditure, the models' ability to correctly order conditions by their energy cost is crucial for applications like optimizing assistive devices. The models performed well in this regard, especially for inclined and loaded walking.
Key Results
- Accuracy:
- For assisted walking, the mean absolute percent errors were:
- Novel Condition: 4.4%
* Novel Subject: 8.0%
* Both-Novel: 8.1%
- For inclined and loaded walking, the mean absolute percent errors were:
- Novel Condition: 6.1% * Novel Subject: 9.7% * Both-Novel: 11.7%
- Ordering Accuracy: The models correctly ordered the magnitude of energy expenditure across conditions:
- Assisted Walking: 63% (average percentage of correctly ordered conditions)
- Incline and Loaded Walking: 87% (average percentage of correctly ordered conditions)
- Model Performance: Neural networks generally performed slightly better than linear regression, especially in more complex scenarios. Recurrent neural networks for fixed time intervals showed comparable performance to per-gait cycle models.
- Input Simplification: Restricting inputs to only vertical ground reaction forces and EMG signals minimally impacted accuracy, suggesting that even simplified sensor setups could be effective.
Limitations and Future Work
- Lab-Based Data: The study used data collected in a controlled lab environment. Future work should investigate the efficacy of these methods using completely wearable sensors in real-world settings, which may introduce more noise and reduce performance.
- Dataset Size and Diversity: While the datasets were representative, a truly generalizable model would require significantly more data across a wider range of conditions and subjects.
- Hyperparameter Tuning: The validation set used for hyperparameter tuning was relatively small. Future work could explore more robust validation strategies.
- Absolute vs. Relative Accuracy: The models are generally better at capturing relative changes in energy expenditure than absolute values, which is important to consider for practical applications.
- Feature Engineering: For linear regression models, further feature engineering or selection could potentially improve performance.
The authors provide their models, code, and datasets, encouraging further research and development in this area.
Summary updated