Validated Caloric Expenditure Estimation using a Single BodyWorn Sensor Jonathan Lester Carl Hartung Laura Pina Ryan Libby Gaetano Borriello Glen Duncan Ubicomp 2009 Presenter Laura Pina CSE 291 Why Do We Need a System to Estimate Calorie Expenditure In 2007 35 of US adults were considered overweight by the US Center for Disease Control Serious Health Illnesses caused by an Overweight hart diseases heart stroke some forms of cancer type 2 diabetes and hypertension Health care costs exceeds 100 billion dollars Why Do We Need a System to Estimate Calorie Expenditure Research has shown people overestimate the calories they expend and underestimate the calories they consume Thus assisting in monitoring calories consumed vs expended would be informational for individuals as well as for their doctors Have you tried to maintain two diaries for 1 2 foods consumed throughout the day type and duration of physical activity throughout the day Have you tried to maintain two diaries for 1 2 foods consumed throughout the day type and duration of physical activity throughout the day VERY DIFFICULT TO BE ACCURATE AND PRECISE Solution Create a system which can automatically track and provide user feedback about their energy balance Two components to energy balance 1 Keeping track of foods consumed throughout the day 2 A system able to compute caloric expenditure Today we will focus on 2 A system able to compute caloric expenditure Order of Presentation Motivation Why Do We Need a System to Estimate Calorie Expenditure Experiment Details MSP Experiment Setup iMote MSP Ground Truth Standard Medical Measure Equation for Caloric Expenditure Activity Inference Results Future Work Experiment Details Laboratory and a Field Experiment Each subject was asked to perform one field session one lab session a third randomly assigned to be either one of the above Each subject wore an Intel mobile Sensing Platform MSP and a VO2 mask data collecting system Total of 51 subjects of varying age ethnicity sex and body type MSP Experiment Setup Users wore the MSP on their waist Sensors used from the MSP accelerometer barometer Data from these sensors provided the data to compute caloric expenditure iMote MSP What sensors were used Accelerometer sampled at 512 and features computed at 4Hz Barometer sampled at 15 Hz Ground Truth VO2 allows to measure energy consumption A metabolic measurement system was used to compare the energy consumption The system measures the flow O 2 content and CO2 context inspired and expired gases This allows to estimate how much O 2 the body has consumed and CO2 produced and then infer metabolic rate VO2 is the most readily available and reliable measure of caloric expenditure Standard Medical Measure Equation for Caloric Expenditure Weight is used for resting metabolic rate Height and Gender used to compute stride length What do we need to calculate calories expended Weight Type of physical activity Speed the user is performing the activity Grade the user is performing the activity What do we need to calculate calories expended Weight Type of physical activity Speed the user is performing the activity Grade the user is performing the activity Order of Presentation Motivation Why Do We Need a System to Estimate Calorie Expenditure Experiment Details MSP Experiment Setup iMote MSP Ground Truth Standard Medical Measure Equation for Caloric Expenditure Activity Inference Results Future Work Activity Inference Two separate Na ve Bayes classifiers were used One was trained and used on Lab Data and another was trained and used on field data The trained classifier focused on identifying the type of activities related to the equation calculating caloric expenditure Resting Walking and Running The features used in the classifier are based are based on the 3D acceleration vector the sum of on second FFTs using frequency bins of 2 4Hz and 1 10Hz variance standard deviation range and a step speed estimate from a step counter Step Counter implemented to find the speed at which the activity is being performed Barometric Pressure sensor data is the simplest and lowest power consumption method of estimating the grade of the activity Step Counter Once the activity has been identified need to calculate speed and adapt to user s walking characteristics such as gate A modified version of Pan Tompkins method using adaptive FFT energy based filter This energy threshold allows to robustly use the same method for a variety of walking conditions users and sensor position Accuracy At Walking Speeds 91 with std dev of 8 1 At Running Speeds 90 with std dev of 9 8 Example of Step Detection Output Example accelerometer trace with the magnitude of the accelerometer shown in solid line at the bottom extracted footfall peaks are marked with X s and the estimated speed from the steps is shown as a dotted line During the first half of the trace the subject was walking at around 3MPH and then began running at 4 5MPH near the 10 44 mark Grade Computation 1 2 3 4 Barometric Pressure Barometric Pressure and GPS GPS GPS and Geographical Information System GIS Accuracy Original Estimate 86 29 Barometer Only 95 88 Barometer GPS 95 25 GPS USGS DEM 95 65 GPS LIDAR Scans 88 61 GPS Only 92 98 Improvement 9 59 8 96 9 36 2 32 6 69 Why do we need to compute Slope 1 2 Improves over all accuracy of the system medical equation Reward users for their strenuous activities More calories are burned when one is being physically active on a slope Example of Estimated Grade of Activity Top A plot of the estimated altitudes Bottom A plot of the estimated grades How were the Results obtained Analysis of data was done by 1 Aligning sensor data with ground truth and VO2 data 2 Running data through the classifier 3 Once an interval of activity was identified the corresponding equation was used Step speeds were calculated from the step detector Weight is used for resting metabolic rate R Height and Gender used to compute stride length Lab Results Accuracy 89 52 std dev 7 25 Top Example field data with the ground truth labels as provided by the observers Bottom Re labeled activities listed as either sitting still or walking moving activities The darkened portions of the accelerometer trace are those inferred by the activity inference as being walking Field Results Accuracy 79 8 std dev 7 25 Accelerometer magnitude solid blue waveform versus the smoothed ground truth VVO2 dotted line reported by the metabolic cart for an example laboratory data trace The
View Full Document
Unlocking...