Ambulatory Energy Expenditure Estimation: A Machine Learning Approach

AAAI 2010

TL;DR

This paper introduces "Jog Falls," a machine learning system for accurately estimating energy expenditure during daily activities. By fusing data from accelerometers (worn on the chest and hip) and a heart rate monitor, Jog Falls intelligently switches between sensors to overcome individual sensor limitations and achieve more accurate results than existing off-the-shelf solutions.

What Problem Does This Paper Address?

The paper tackles the growing problem of obesity and its associated health issues like diabetes and heart disease. A key factor in managing weight is understanding energy expenditure (EE) and balancing it with energy intake. While many wearable devices exist to estimate EE, they often suffer from limitations such as inaccuracy, poor system integration, and susceptibility to noise or data loss. For instance, pedometers only track steps, while some advanced devices can be misled by vibrations (e.g., in a car) or fail to account for upper body movements. This paper aims to provide a more robust and accurate solution for ambulatory EE estimation that addresses these shortcomings.

Methods

The core of the Jog Falls system is a Bayesian Network (BayesNet) that acts as a "fusion algorithm." This algorithm intelligently selects the most appropriate sensor(s) for estimating EE based on the current context.

Here's a breakdown of the key components and techniques:

  • Multi-Sensor Fusion: The system utilizes data from three main sources:
  • Chest Sensor: Combines an ECG heart rate (HR) monitor and a 3-axis accelerometer (CA) to estimate upper body movement intensity (in Metabolic Equivalents, or METs, referred to as CA_MET).
  • Hip Sensor: A multi-sensor platform with a 3-axis accelerometer (HA) to identify user activity (sitting, walking, running) and estimate walking speed.
  • Data Aggregator (Cell Phone): Collects data from the sensors, aggregates it, and feeds it into the Data Fusion algorithm.
  • BayesNet for Sensor Validity: The BayesNet is trained to infer the "validity" of each sensor. It uses features extracted from the sensor data (e.g., Delta Heart Rate, Lower Body Speed, Upper Body MET) as evidence to determine the conditional probability of a sensor being valid. This allows the algorithm to:
  • Identify Sensor Noise: For example, it can detect when vehicle vibrations might be misleading the accelerometer and switch to the heart rate sensor.
  • Handle Data Loss: If one sensor fails, the BayesNet can still operate by integrating over the missing variable, providing graceful degradation.
  • Compensate for Individual Sensor Weaknesses: It leverages the strengths of each sensor. For instance, it uses accelerometers for sedentary activities and heart rate for higher intensity activities. It also accounts for situations where heart rate might overestimate EE during recovery phases.
  • Individual Calibration: The system incorporates individual calibration for heart rate to account for variations in fitness levels. Participants perform controlled walking at different speeds to build a personalized HR vs. MET calibration table.
  • Feature Engineering: Several features are extracted from the sensor data, including:
  • Delta Heart Rate
    • Heart Rate Forward Difference
    • Lower Body Speed
    • Upper Body MET
    • Mean Heart Rate
    • Heart Rate Mean Difference

Datasets

  • Calorimeter Experiment Data: A lab experiment was conducted using a calorimeter (considered a gold standard for EE measurement) with four subjects performing five different activities (slow walking, brisk walking, running, weight lifting, biking). This data was used to train the fusion algorithm and understand sensor limitations.
  • User Study Data: A 15-participant, 63-day user study was conducted in India. This real-world data was used to validate the system's performance in ambulatory conditions.

Key Insights and Ideas

  • Context is King for EE Estimation: No single sensor is perfect for all situations. The key to accurate ambulatory EE estimation lies in dynamically selecting the best sensor(s) based on the user's current activity and environmental context.
  • Fusion Mitigates Individual Sensor Failures: The BayesNet's ability to infer sensor validity and gracefully degrade when a sensor fails is a critical innovation. This makes the system more robust than solutions that rely on a single sensor or simple averaging.
  • Vehicle Vibration is a Major Challenge: Accelerometers are highly susceptible to vibration noise (e.g., from vehicles), leading to overestimation of EE. The fusion algorithm's ability to detect this and switch to heart rate is a significant improvement.
  • Heart Rate and Accelerometer Complement Each Other: Heart rate tends to overestimate EE during recovery from intense exercise, while accelerometers can underestimate exertion. The fusion algorithm intelligently switches between them to balance these tendencies.
  • Psychological/Emotional Factors Impact HR: The paper acknowledges that factors like anxiety can elevate heart rate, leading to inflated EE estimates. The fusion algorithm attempts to mitigate this by considering movement data alongside HR.

Key Results

  • Improved Accuracy: The fusion algorithm achieved an estimated EE within +5% of calorimeter measurements, significantly outperforming individual sensors which showed errors of +22% (heart rate) and -10% (accelerometer) in controlled experiments.
  • Comparative Performance: In a comparison with BodyBugg, the Jog Falls algorithm closely tracked EE for most activities, with an aggregate difference of -5% compared to BodyBugg's -6% (and BodyBugg overestimated EE in vehicle travel).
  • User Study Performance: The user study showed that the fusion algorithm's sensor choices mirrored those found effective in calorimeter tests. Cumulative EE estimates from the fusion algorithm were higher than those from individual accelerometer or heart rate sensors, which underestimated EE by 15-20% due to factors like intermittent data loss.
  • Graceful Degradation: The system demonstrated the ability to operate even with missing sensor data, with an average run time of 13 hours a day and only 30% of the time experiencing issues with one sensor.

Limitations and Future Work

  • Lower Body Speed Calibration: The mapping of lower body speed to MET values was based on a general compendium ([Ainsworth 2000]) and was not individually calibrated for each user. The authors intend to address this in future revisions.
  • Training Data Generality: The training dataset was not extensive enough to claim complete generality of the solution.
  • Future Work: The authors plan to further improve the solution by addressing the limitations mentioned above, particularly individual calibration for lower body speed.

Summary updated

BibTeX
@inproceedings{2010ambulatory,
  title = {{Ambulatory Energy Expenditure Estimation: A Machine Learning Approach}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
}
Figure