Accelerometry-based Energy Expenditure Estimation During Activities of Daily Living: A Comparison Among Different Accelerometer Compositions

Technical Summary: Accelerometry-based Energy Expenditure Estimation During Activities of Daily Living

Problem Statement
Physical Activity Energy Expenditure (PAEE) is a critical parameter for obesity control and health monitoring, exhibiting high variability both between individuals and within a single day. While PAEE can be accurately measured via indirect calorimetry (breath-by-breath respiratory data), this method lacks portability for daily life monitoring. Consequently, there is a need for wearable sensor-based tools, specifically accelerometers, to estimate PAEE. A central challenge in this domain is determining the optimal sensor location. While the body's center of mass (COM), typically near the lower back or pelvis, is theoretically ideal for capturing whole-body movement, the wrist has become a popular location due to the prevalence of consumer wrist-worn devices (e.g., Apple Watch, Fitbit). However, the efficacy of wrist-based sensors in estimating PAEE compared to COM-based configurations remains under-investigated, particularly regarding whether single sensors suffice or if multi-sensor compositions are required to approximate COM acceleration.

Methodology
The study evaluated and compared four distinct accelerometer compositions for PAEE estimation during Activities of Daily Living (ADL):

1. COM-based settings:
   • Pelvis-acc: A single accelerometer on the pelvis.
   • 3-acc: One pelvis accelerometer combined with two thigh accelerometers.
2. Wrist-based settings:
   • l-wrist-acc: A single accelerometer on the left wrist.
   • r-wrist-acc: A single accelerometer on the right wrist.

Data Collection:
Ten participants (30% female) performed a randomized sequence of ADLs, including sitting, standing, laptop work, household chores (mopping, dishwashing), stair climbing, and treadmill/cycling exercises. Respiratory data (VO2 and VCO2) collected via a COSMED K5 system served as the ground truth for PAEE. Five Movella Xsens DOT sensors (30 Hz) recorded 3-axis acceleration at the pelvis, both thighs, and both wrists. Data preprocessing involved removing gravitational acceleration, low-pass filtering (6 Hz), and resampling to 1 Hz. PAEE was derived by subtracting Resting Metabolic Rate (RMR) from Total Energy Expenditure (TEE).

Models:
Two established estimation methods were implemented on the dataset:

1. Linear Regression (LR): Utilized the feature $IAA_{tot}$ (sum of integrated absolute acceleration across three axes) as input. For the 3-acc setting, three independent $IAA_{tot}$ variables were used.
2. CNN-LSTM: A deep learning model taking pre-processed 3-axis acceleration data as input. For single sensors, the input matrix was 3x30 (3 axes, 30-second window); for the 3-acc setting, the feature dimension was concatenated to 9x30.

Both models performed 1-second horizon forecasting using a 30-second sliding window. Performance was evaluated using Leave-One-Subject-Out cross-validation, Normalized Root-Mean-Squared Error (NRMSE), and R-squared ($R^2$) values. Statistical significance was assessed via repeated measures ANOVA and paired t-tests with Bonferroni correction.

Key Results

• Model Performance: Both the LR and CNN-LSTM models yielded their best results with the 3-acc (COM-based) setting. The CNN-LSTM model achieved the highest $R^2$ of 0.53 (NRMSE = 0.64), while the LR model achieved an $R^2$ of 0.41 (NRMSE = 0.72).
• COM vs. Wrist: The two COM-based settings (3-acc and pelvis-acc) significantly outperformed the wrist-based settings ($p < 0.05$). The wrist-based settings demonstrated almost no predictive power, with $R^2$ values close to 0 (ranging from -0.02 to 0.06) and high NRMSE values (~0.95).
• Sensor Composition: There was no significant difference between the single pelvis sensor (pelvis-acc) and the three-sensor composition (3-acc) for either model ($p = 0.278$ for $R^2$), suggesting a single sensor near the COM is sufficient to represent COM acceleration for this task.
• Wrist Symmetry: No significant difference was found between the left and right wrist sensors ($p = 0.329$).
• Model Comparison: No significant difference was found between the LR and CNN-LSTM models in terms of NRMSE or $R^2$ ($p > 0.05$), indicating that the linear feature $IAA_{tot}$ is comparable in predictive power to the more complex CNN-LSTM architecture for this specific dataset.

Significance and Claims
The paper concludes that for accurate PAEE estimation during ADL, COM-based modeling approaches are superior to wrist-based approaches. The authors claim that wrist-worn accelerometers, when used in isolation, fail to reflect COM acceleration effectively, resulting in negligible predictive power for PAEE. This challenges the usability claims of commercial wrist-worn devices that rely solely on acceleration for energy expenditure tracking.

The study suggests that a single sensor placed close to the COM (e.g., pelvis) is sufficient to capture the necessary movement dynamics, as adding thigh sensors did not significantly improve performance over the pelvis-only setup. The authors maintain a modest stance, noting that while wrist-based estimation is currently limited, performance might improve with additional sensor modalities (e.g., heart rate) or if methods are developed to better map wrist acceleration to COM acceleration. They recommend that future research focus on validating these findings with larger datasets and exploring other sensor locations (e.g., chest, upper arm) that may better approximate COM acceleration.