Estimating Energy Expenditure from Activities

Because both voluntary activities and involuntary activities are related to energy expenditure, we can potentially use wearable inertial motion sensors to monitor activities and estimate energy expenditure (EE). On top of that, thanks to the invention of wearable ECG sensors and pulse-oximeters to track heart rates (HR), we can also combine the inertial motion sensors and HR sensors to provide more information about one’s EE during activities.

Of course, machine learning algorithms can be employed to do the estimation. But which one? How do we go from activities to energy expenditure?

The classic approach in the wearable sensor community goes like this:

  1. Collect motion sensor data, recognize activity patterns from the motion data.
  2. Convert activity types to something called Metabolic Equivalents. This is done by reading the Metabolic Equivalent Table.
  3. Calculate accumulated energy expenditure based on the duration of each activity.

Several annoying problems are associated with this approach:

  1. Activity recognition (AR) is a supervised learning approach. It requires the activities to be labeled in the training data.  Labeling means researchers need to do record the timestamps of the beginning and the end of each activity, and activity types.
  2. The AR models built in-lab can be less predictable in real-world. Since the researchers cannot test every type of activity in-lab and humans are so creative at coming up with activity patterns, AR models don’t scale well to the real-world. Categorizing an activity wrongly could further affect the estimation of energy expenditure.
  3. More accurate models usually need more sources of information, typically this means more sensors to wear. Researchers can also tweak the machine learning algorithms to make them more robust, the accuracy boost is often incremental.
  4. The worst bit. Since this approach does not capture the variance in energy expenditure during activities between different subjects and different intensity levels, it’s not very accurate.

To overcome Problem 4, researchers adopt regression method along with the activity recognition, however, it doesn’t address Problem 1, 2 and 3.

Can we use one sensor, skip the AR step, and still provide good accurate energy expenditure estimation?

My answer is documented in this paper. In this project, I used a wearable ECG patch developed by imec. The patch integrated a wearable motion sensor and an ECG sensor. To skip the labeling step, one can employ unsupervised learning, such as clustering. After clustering, activities of similar intensity will be roughly grouped together. Within each cluster, we can then regress the energy expenditure on the sensor data.

 

 

One big caveat of using wearable sensors is that they are only accurate at estimating EE during activities. The main component of daily EE is from resting EE, a.k.a. EE of the subject during the resting period, unfortunately, cannot be measured by the wearable sensors. How to model resting EE is still a research topic.