Unsupervised Activity Clustering to Estimate Energy Expenditure with a Single Sensor

Shanshan Chen, John Lach

Marco Altini, Julien PendersOliver Amft

2

2H218O

Existing Solutions

BSN?

Research on Energy Expenditure (EE) Estimation with BSN

• Detailed Activity Recognition (AR) + Metabolic Equivalents (METs)• Annotation labeling work at the development stage• Lots of sensors to wear for the users• Lack of accuracy due to static number of METs

• Detailed AR + regression• Labeling work at the development stage• More inertial sensors needed for better recognition accuracy

• Detailed AR Grouped AR + regression• Reduced number of sensors – ECG + Accelerometer• Reduced challenges in high accuracy recognition

• Data-driven clustering + regression• Bypass activity recognition• No labeling at the development stage

3

Proposed method• Focus on accurate EE estimation, not AR• Clustering based on motion and heart rate, not activities

• Data-driven clustering• Apply regression model based on data cluster

• Unsupervised learning• No need to label activities during development stage

• EE accuracy independent of AR accuracy

4

Features from Data

Clustering

Group 1

Group 2

Group N

Model1

Model 2

Model N

Experiment Setup

5

▪ Single sensor node data (acceleration + heart rate) and validation data (circulatory calorimeter) collection

▪ 10 subjects of various BMI

▪ 52 types of activities (sedentary activities and physical exercises)

Feature Extraction -- Preprocessing▪ Heart rate

▪ Removing the motion artifact▪ Count peaks every 15 seconds▪ Extract heart rate above rest

▪ Acceleration features extraction▪ 4 seconds time window▪ 18 features extracted in total

6

Feature Extraction Machine Learning

Framework of Machine Learning

7

Feature Selection(LASSO)

Multiple Linear

Regression

Dimension Reduction

19 Features

Model Comparison• Proposed model

• Apply different regression models to different data clusters • Single multiple-linear regression model

• Also activity-oblivious• Single regression model

• AR-based model (Grouped AR + Regression)• Perfectly separated based on known activity labels• Non-ideally separated based on AR algorithms

8

Regression Results

9

▪ Proposed model is better than the single regression model

▪ With perfect labeling, activity specific model is the best

▪ However, accuracy of AR based method drop quickly when misclassification happens

Future Work• Explore other unsupervised learning techniques• Study interpretations of clusters

• Histogram of activities inside each cluster

• Real-time implementation• Monitoring intensive activities only to save battery

• Greater subject diversity• Combine with emerging energy intake techniques

10

Conclusion• Data-driven clustering for EE estimation

• One light-weight sensor patch, easy for the users to wear• No labeling of activities at the development stage• Final estimation accuracy does not depend on accuracy of AR

• Improve linear regression model and AR based clustering• Drawback:

• Does not track activities – orthogonal problem of accurate energy expenditure estimation

11

THANKS!

?

Histogram of Activities in Clusters

13

å

Clustering Results

14

Training set clustering Testing set clustering

Physical Activities Comparison

15

▪ Physical activities are more interesting to monitor instead of the sedentary ones

▪ The proposed model achieves almost as good accuracy as activity specific model