department of computer and electrical engineering a study of time-based features and regularity of...

43
partment of Computer and Electrical Enginee A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity Periods During Free-Living MS Defense Exam Jose Luis Reyes Dr. Adam Hoover (chair) Dr. Eric Muth Dr. Richard Groff April 24, 2014

Upload: samuel-cunningham

Post on 02-Jan-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Department of Computer and Electrical Engineering

A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Periods During Free-Living

MS Defense ExamJose Luis Reyes

Dr. Adam Hoover (chair)Dr. Eric Muth

Dr. Richard Groff

April 24, 2014

Page 2: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

OutlineMotivation and BackgroundDesign and MethodsResultsConclusion

Page 3: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Obesity• Common

– 34% of U.S. population are obese [Centers for Disease Control and Prevention]

• Serious– 5th leading risk for global deaths [WHO, 2014]– Heart disease, stroke, type 2 diabetes, and certain types

of cancer [Centers for Disease Control and Prevention]• Costly

– In 2008, annual medical cost was $147 billion in the U.S. [Centers for Disease Control and Prevention]

– In 2008, medical cost was $1,429 higher than of those of normal weight. [Centers for Disease Control and Prevention]

Page 4: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Obesity treatmentsDietary changesExercise and activityBehavior changesWeight-loss medicationWeight-loss surgeryLimit energy intake (EI)*

Balancing EI and EE (energy expenditure)

Page 5: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Monitoring EIMost widely used tools

Food diary24-hour recallFood frequency questionnaire

Technology-based toolsCamera [Martin et al., 2009]Wearable sensors [Amft et al., 2008]

Page 6: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Bite Counter

Watch-like deviceWrist motion trackingAccelerometer and gyroscope

Page 7: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Previous work

Goal: Detection of eating activity periodsBased on accelerometer (AccX, AccY, AccZ)

and gyroscope (Yaw, Pitch, Roll) readingsData segmentationClassification of eating activity (EA) and non-

eating activity (non-EA) periods based on features

Overall accuracy obtained was 81%

Page 8: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

NoveltyPrevious work considered only sensor-based

featuresWe consider the time component

Time since last eating activityCumulative eating time

Periodicity of manipulation over timeRegularity of manipulation

Page 9: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Design and methodsOverview of algorithmData collectionNew features

Regularity of manipulationTime since last EACumulative eating time

Evaluation metrics

Page 10: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Overview of algorithm (Dong et al., 2013) •Data smoothing

- Gaussian kernel

Page 11: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Overview of algorithm

Sum of acceleration,

Page 12: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Overview of algorithmData segmentationPeak detection

Sum of accelerationHysteresis

threshold

Page 13: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Overview of algorithmFeatures

Manipulation

Linear acceleration

Wrist roll motion

Regularity of roll

Page 14: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Overview of algorithmNaive Bayes Classifier

Assign most probable class, ci in C

Given features f1,f2, …, fN

Feature probability

Page 15: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Data collectionCollected using iPhone 4

Programmable , large amount of memory, accelerometer and gyroscope

Recorded at 15Hz2 sets of data

Set 1: 20 recordingsSet 2: 23 recordings

A total of 449 hours of dataData training

5 minute non-EA segmentsFull segments for EA

Page 16: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Current work

Motivation: improve previous accuracy of 81%

Introduction of 3 new features:Regularity of manipulationTime since last EACumulative eating time

Page 17: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

FeaturesFeature 1, regularity of manipulation

Regularity of peaks around 4000-5000 (deg/s)/G

Peaks every 10 – 30 seconds?

EA manipulation segment Non-EA manipulation segment

Page 18: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Regularity of manipulationSmooth manipulation data (N = 225, R =

37.6)Compute FFTCompute:

Units: (deg/s3)/G

Page 19: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Regularity of manipulationCalculate for each segment in dataDistribution statistics can be used for Bayes classifier

29>>

Distributions (set 1)

Page 20: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Regularity of manipulation

Distributions (set 2)

34>>

Page 21: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Features

Feature 2, time since last eating activityTime componentAfter a person eats, very unlikely to eat again

immediatelyProbability starts increasing as time passes

Page 22: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Time since last EALet tlast = end time of last segments classified

as EALet t = middle of time of unknown segment

currently being classifiedThen,

Page 23: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Time since last EABayes classifier requires probability

distributions for both EA and non-EAIt is possible to calculate time between mealsNonsensical for opposite class

Time since last non-EA?1 – p(f|EA)

Page 24: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Time since last EA Compute cumulative distribution function (CDF) of time since last

EA. p(f|EA) = CDF, p(f|nonEA) = 1 - CDF

CDF for time since last EA

(set 2)

Page 25: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Features

Feature 3, cumulative eating timeTime componentPeople spend a certain amount of time eating

and drinking in a day(Around 1.1 hrs. according to Dept. of Labor Statistics )

Page 26: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Cumulative eating timeAt time t, cumulative eating time:

Distribution of times involving non events are nonsensical

Compute CDF for each recording and average in each data set

Page 27: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Cumulative eating time

CDF for cumulative eating time (set 2)

Page 28: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Cumulative eating timep(f|EA) =

σ2cdf, μcdf from average CDF

p(f|nonEA) = 1 – p(f|EA)

Page 29: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Evaluation metricsOverall accuracy

EA accuracy

Non-EA accuracy

Page 30: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

ResultsPrevious work

Statistics

Accuracy

Page 31: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

ResultsRegularity of manipulation

Statistics

Accuracy

Page 32: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Regularity of manipulation (Results)

Standard deviation relatively large for EA distribution (<<18)

Set 1’s EA distribution non GaussianFFT not completely discriminating between

EAs and non-EAs

Page 33: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Regularity of manipulation (Results)

Smoothed manipulation segment from EA distribution (right tail)

Smoothed manipulation segment from non-EA distribution (left tail)

Page 34: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Regularity of manipulation (Results)Smoothed manipulation segment from EA distribution (middle)

Smoothed manipulation segment from non-EA distribution (middle)

<<20

Page 35: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Regularity of manipulation (Results)

Original data for segment in middle of EA distribution

Original data for segment in middle of non-EA distribution

Page 36: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

ResultsTime since last EA

Statistics

AccuracySet 1 Set 2

Page 37: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Time since last EA (Results)

Original 4 featuresOriginal 4 features + time since last EA

Page 38: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Time since last EA (Results)

Original

Including time since last EA

• FPs are strong inhibitors for immediately subsequent data

Page 39: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

ResultsCumulative eating time

Statistics

AccuracySet 1 Set 2

Page 40: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Cumulative eating time (Results)

Original 4 featuresOriginal 4 features + cumulative eating time

Page 41: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Cumulative eating time (Results)

Original

Including cumulative eating time

• FPs are strong inhibitors for immediately subsequent data

Page 42: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Conclusion

FFT not discriminating between EAs and non-EAs completely

Time-based features act as clocksFuture work

Explore regularity of manipulation using non-sinusoidal transform

Explore off-line analysis using time-based features so the optimal daily solution can be found (HMMs)

Page 43: Department of Computer and Electrical Engineering A Study of Time-based Features and Regularity of Manipulation to Improve the Detection of Eating Activity

Questions?