2013 IEEE Point-of-Care Healthcare Technologies (PHT) Bangalore, India, 16 - 18 January, 2013

On Determining the Best Physiological Predictors of Activity Intensity Using Phone-Based Sensors

Harshvardhan Vathsangam, E. Todd Schroeder and Gaurav S. Sukhatme

Abstract-Physical inactivity is a leading risk factor in worldwide deaths. This problem has led to the need for new research paradigms investigating the ef1"ect of sedentary behavior on negative health outcomes. Central to this need is the development of objective and Ubiquitous sensors that provide accurate measurements of activity to assist in intervention. Phone-based kinematic sensors, such as accelerometers and gyroscopes, are one such option. Current kinematic sensor models have limited capability in adjusting for inter-personal physiological differences in the maps from movement to activity intensity since they focus on weight and height information. It would be useful to explore what features are the best descriptors for a population. We present a family of regression techniques that incorporate an arbitrary number of physiological features and use this framework to determine the best physiological features to map movement to energy expenditure. We do this for rest, treadmill and overground walking since these are the most common activities for which intervention is necessary. Size­based features, such as height, weight and BMI were the best descriptors for personalization. BMI was the best descriptor for rest and height was the best descriptor for walking. Fitness based features, such as resting energy expenditure and resting heart rate, were the least useful descriptors, particularly for walking.

I. INTRODUCTION

Physical inactivity is the fourth leading risk factor for causes of deaths worldwide [1] and plays an important role in disease-related causes of heart disease, malignant neoplasm, cerebrovascular disease and diabetes mellitus [2]. The in­creasing evidence for negative effects of sedentary behavior has necessitated new research paradigms to explore the rela­tionships between sedentary behavior with risk biomarkers and health outcomes [3]. A suggested intervention is the promotion of low intensity physical activity, such as walking or gardening in place of sedentary time [4].

Central to this research is the requirement of objective and ubiquitous trackers of activity levels to provide objective feedback and assist in point-of-care intervention measures. Kinematic sensors such as accelerometers have played an important role in providing such measures [5, 6]. Current kinematic sensor-based activity research is limited in that the algorithms used to represent movement on these sensors are not supported by meaningful modeling[7]. One particular issue is that while two individuals might make the same amount of movement, the actual activity toll on their body will differ. The model from movement to activity intensity will depend on physiological descriptors such as height, weight, sex, fitness, etc[5]. Current kinematic sensor-based

H. Vathsangam and G. S. Sukhatme are with the Dept. of Computer Science, Univ. of Southern California, Los Angeles, CA 90089, USA.

E.T. Schroeder is with the Division of Biokinesiology and Physical Therapy, Univ. of Southern California, Los Angeles, CA 90089, USA.

This work was supported in part by NSF (CCR-0120778) as part of the Center for Embedded Network Sensing (CENS). Support for H. Vathsangam was provided by the USC Annenberg Doctoral fellowship.

models do not extensively model interpersonal differences, instead relying on normalized measures of activity intensity such as energy expenditure scaled by a power of the weight or height [8]. What is needed is a framework that incorporates as many descriptors as necessary and allows researchers to pick the best descriptors for their population from data. A

related question is which descriptors are more important and whether one can detennine them from the population itself.

In this paper, we describe a set of regression-based nor­malization techniques that map cellphone accelerometer and gyroscope data to activity intensity for three activities: rest, treadmill and overground walking. These activities were chosen because rest and walking represent the most common activities in our day and accurately characterizing their in­tensity would provide valuable tools in quantifying sedentary lifestyles. We use these models as tools to provide an evalu­ation of different physiological descriptors according to their ability to personalize energy expenditure models as measured by lowering prediction error. We use energy expenditure as a representation of activity intensity. A Bayesian Regression model [9] is used to map physiological descriptors to Resting Metabolic Rate (RMR). A hierarchical model [10] is used to obtain a normalized map from movement to Total Energy Expenditure (TEE). We use statistical methods to rank phys­iological descriptors by their relative importance in predicting activity intensity. The novelties of this paper include using previous techniques[ 10] to validate best descriptors on a larger population and validation for overground walking.

II. ALGORITHMS

A. Resting Metabolic Rate Estimation

We adopt a Bayesian approach through Bayesian Linear Regression (BLR). The approach is similar to our earlier work in predicting energy expenditure [9]. In the context of energy prediction, for a given person p, with input physiolog­

ical information PhY2 = {Physnp} �::l and target energy

values Y p = {Ynp} n::l' we assume a linear dependence between the physiological descriptors and energy expenditure for each person:

Ynp w� Physn.p + E, E '" N (0,0";2) ,

where E is a noise parameter and w p = (wop, ... , W M -lp) T are the model weights with a Gaussian prior, p( w p) =

N(wp;O,o:-ll). Training the model is equivalent to learn­ing the weights wp and the noise parameters O"p. The optimal prediction for a new data point is given by the predictive distribution:

p(y*pIPhys*p, Yp, 0:, O"p)

0"7v (Phys*p) p mNp S-l Np

0";2 + Phys*p TSNpPhys*p,

O";SNPhys�Yp,

0:1 + O";Phys�Physp"

978-1-4673-2767-1/13/$31.00 ©2013 IEEE 140

The model parameters are learned using the EM algorithm.

B. Total Energy Expenditure Estimation

We use a Hierarchical Linear Model (HLM) to capture inter-person and intra-person variations in mapping move­ment to energy expenditure. Consider a test population consisting of P participants. For each participant p, let there be Np data points collected, consisting of the energy values Yp = {YIp' Y2p' ... , YNp} and D-dimensional fea­ture inputs Xp = {XIp,X2p' ... ,XNp}. We denote Y =

{YI,Y2, ... ,Yp} and X = {XI,X2, ... ,Xp} to be complete set of training data for all participants. Let each par­ticipant have physiological descriptors determined by Physp and the complete set be PHY = {PhYSp}:=I. These include parameters such as height, weight and age (with a constant term for bias). We model top-down dependence of wp on each participant's physiological descriptors Physp through a universal weight parameter k. Each wp in turn influences energy predictions Ynp for an input xnp• Each output energy value, Ynp is linearly dependent on input xnp• This can be expressed as:

Yp rv N (Yp; w�Xp, 0";;-21) .

.We assume that each per-person model parameter wp has a linear dependence on k and Physp:

wp rv N (wp; kTphysp, a-II) .

Both wp and k are hidden variables which need to be estimated from data. We denote W = {Wp}:=1

Training the multilevel model is equivalent to learning individual wp's, the overall parameter k as well as the noise parameters {O"P}:=I' a and 0". The HLM combines P personal regression models in two ways. First, the lo­cal regression coefficients W p determine energy values for each person. Second, the different coefficients are connected through the population-level model parameter k. Intuitively, the HLM captures the inherent similarity in walking across different people while accounting for individual walking styles and energy consumption. This allows us to instantiate a model for a new person using only top-level physiological descriptors without having any training data from them. This model is trained using an EM-like algorithm [10].

e. Physiological Descriptors

The physiological descriptors used to describe a person during rest were height, weight, BMI, gender and resting heart rate. The descriptors used to describe a person during walking were height, weight, BMI, gender and resting heart rate and resting energy expenditure. Height, weight were indicators of the size of a person. BMI was an indicator of total body fat. Resting heart rate and resting energy expenditure were indicators of fitness.

III. METHODS A. Data Collection

A total of 43 participants (27 male, 16 female) participated in the study. All participants signed informed consent forms

1.95

1.9

1.85

I!! 1.8 CD a:; E- 1.75 1: OJ

'a; 1.7 I

1.65

1.6 •

1.55

1.5

Normal Overweight

_Obese • Males . Females

• •

•

• •

• •

•

50 60 70 80 Weight (Kgs)

90 100 110

Figure 2: Statistics of test population. Males, Average Height: 1.76 ± .07m , Average Weight: 74.08 ± 9.9kg. Females, Average Height: 1.65±.05m, Average Weight: 58.40±8.2kg. The colored regions represent healthy (Yellow, BMI<25), overweight (Light orange, BM1>25) and obese individuals (Dark orange, BMI>30).

measuring tape. Each participant's weight was also measured with the EatSmart Precision Digital weighing scale. Figure 2 describes participant statistics.

Energy expenditure was measured using the Oxycon™

Mobile Metabolic unit from Carefusion. The metabolic cart reports participant V02 and VC02 and derived calorie data at the frequency of every breath. The cart was worn as a backpack fitted to the comfort of the participant. Figures Ia and Ib illustrate a typical recording setup.

To record movement, each participant wore a Galaxy Nexus S phone running Android 2.3.3 on the right iliac crest with a belt holder to record movement. Inertial data was recorded with a custom-built smartphone app - Movement Trackr [11]. The app records triaxial accelerometer, triaxial gyroscope and GPS data (when available) at a set sampling rate. For the purposes of this study, the accelerometer settings were set at "Fastest" (�50 Hz) and the gyroscope settings were set at "Game" (�100 Hz). All phone data were stored locally and synchronized in post-processing.

Data collection was carried out in two sessions - indoor and outdoor. In the indoor session, each participant was asked to sit still and meditate for five minutes while basal metabolic rate and resting heart rate data were collected at the frequency of every breath. Each participant then walked on a treadmill at three speeds - 2.5, 3.0 and 3.5 mph for five minutes per speed with two minutes of settling time for each speed. In the outdoor session, each participant was asked to walk on the university athletic track at a self-selected speed for approximately 20 min. Figure Ic illustrates the typical path traversed during the outdoor section of the study.

B. Signal Processing and the study was approved by the Institutional Review Board of the University of Southern California. Before par­ticipation, each participant's height was measured with wall­based height charts and their leg length was measured with a 5

For treadmill walking, for each person, each of the minute streams from the phone were divided into 10

141

(a) Indoor measurement (b) Outdoor measurement (e) Example GPS path

Figure 1: Illustration of data recording procedures for the experiment. Triaxial acceleration and rotational rates were recorded indoors. GPS was recorded outdoors. Time-synced energy and heart rate data were recorded using the metabolic cart.

second intervals or epochs. Within each epoch, six nor­malized Fourier transforms were calculated from triaxial accelerometer and gyroscope data and concatenated. The Fourier coefficients corresponding to frequencies greater than 10 Hz were discarded. These were used as descriptors of movement. These were then synced with mean heart rate and mean total energy expenditure (kcal/min) values for that epoch. Walking data for each user was thus a set of epochs each containing 6-dimensional Fourier transforms and the mean heart rate and total energy expenditure for that epoch. These represent per-user data while walking at three different speeds. The same process was repeated for overground walking.

IV. RESULTS

A. Ranking Methodology

We used a one-of-K testing scheme when evaluating al­gorithms. Given P participants, various regression models were trained on data from P - 1 participants. BLR models were used for rest and HLM models were used for tread­mill and overground walking. In each training instance, all possible physiological feature combinations were tried. For each feature combination, the resultant model was used to predict energy expenditure for the pthparticipant. Given a set of predictions, the normalized root mean squared error was calculated against the ground truth energy expenditure of the pthparticipant for the same data points. This was repeated for each participant in turn and the results were averaged across participants to obtain a feature-specific error performance.

These errors were sorted in ascending order. For each error in this order, the descriptor was awarded a score equal to its error rank. For example, if weight and height appeared in the third lowest error, they both received a score of three. This was done for all errors and the average score for each feature was calculated. The intuition behind this scheme is that if the appearance of a particular feature results in the lower errors, it will appear in the beginning of the sorted list

more often. Hence a lower score for a physiological feature implies greater importance in personalizing a person's energy expenditure model.

B. Rest

Figure 3a illustrates the average scores of different phys­iological descriptors for rest. BMI had the lowest ranking followed by height, resting heart rate and weight. The popula­tion in this study did not include any individuals with low fat reserve and high muscle mass. This suggests a relationship between fat content and energy consumption for rest. Sex did not play a large role since women had a smaller weight and height than men and thus this contribution was absorbed in the body size parameters.This suggests having separate models for men and women and reexamining our models.

C. Walking

Figures 3b and 3c illustrate the average scores of differ­ent physiological features treadmill walking and overground walking respectively. In contrast to the rest case, for both overground and treadmill walking, height had the lowest score and weight the second lowest score. The lower score for height could be once again because the sex contribution is absorbed by the height term. The score difference between height and weight was lower in the case of overground walking as compared to treadmill walking. Resting heart rate showed lower scores for overground walking than for treadmill walking. This lower difference in scores for over­ground walking could be because participants chose a self­selected speed, which would typically be one that minimizes energy expenditure per meter traveled. This self-selected speed would depend on height and fitness level. Resting energy expenditure had the highest score indicating that knowing one's metabolic rate at rest is not enough to be able to predict their energy expenditure for dynamic activities.

142

15 .------,------�-------r------�----__.

OJ C

� 10 c CO

a: Q) OJ CO

Q> >

« 5

O L---�----��----�L-----�----��� Weight Height BMI RHR

Physiological Features

(a) Average descriptor scores for rest

60 r---�-----.----.-----r---�-----.--__.

50

Q) (; 40 <..>

(j) Q) 30 � Q) � 20

10

60

50

Q) ..... 40 0 <..>

(j) Q) 30 OJ � Q) > 20

«

10

0

--

Height Weight BMI Sex RHR REE

Different Physiological Features

(b) Average descriptor scores for treadmill walking

0-r-

rl-���+

Height Weight BMI Sex RHR REE

Different Physiological Features

(c) Average descriptor scores for overground walking

Figure 3: Average descriptor scores for rest, treadmill and free walking (lower is better). For rest, BMI was the best predictor, followed by height, resting heart rate (RHR) and weight. For walking, size-based features like height and weight were the best predictors, followed by sex, BMI and RHR. Resting Energy Expenditure (REE) was the worst descriptor.

V. CONCLUSION

Current accelerometer-based models from movement to activity levels do not completely account for interpersonal differences in physiological parameters. We described a set of regression-based normalization techniques that map cell­phone accelerometer and gyroscope data to activity intensity for three activities: rest, treadmill and overground walking. We used a Bayesian Regression model to map physiological features to Resting Metabolic Rate (RMR) and a hierarchi­cal model to map movement to Total Energy Expenditure (TEE). We then ranked physiological features according to their ability to predict energy expenditure as measured by the lowest percentage prediction error. Size based features such as weight and height were the best descriptors for personalization. BMI was the best descriptor for RMR and height was the best descriptor when walking. Fitness-based features such as resting energy expenditure and resting heart rate were the least useful descriptors for predicting TEE from walking.

We aim to expand on this work by developing explicit parametric models detailing the relationship between these physiological variables and energy expenditure. This exten­sion will involve exploring the mass dependent and mass independent models of energy expenditure. We also plan to use non-parametric approaches to model dependence on physiological parameters and examine the relative importance of variables using these models. We plan to use these tech­niques in providing more accurate, ubiquitous measurements for monitoring and intervention of sedentary lifestyles.

REFERENCES

[I] "Global health risks: Mortality and burdern of diesease attributable to selected major risks," World Health Organization, 2009.

[2] H.-C. Kung, D. Hoyert, 1. Xu, and S. Murphy, "Deaths: Final data for 2005," National vital statistics reports, 2008.

[3] N. Owen, G. Healy, C. Matthews, and D. Dunstan, "Too much sitting: The population health science of sedentary behavior," Exercise and Sport Sciences Reviews, vol. 38, pp. 105-113, 2010.

[4] 1. Levine, "Nonexercise activity thermogenesis - liberating the life­force," Journal of Internal Medicine, vol. 262, pp. 273-287, 2007.

[5] N. Butte, U. Ekelund, and K. Westerterp, "Assessing physical activity using wearable monitors: Measures of physical activity," Medicine and Science in Sports and Exercise, vol. 44(1 Suppl 1), pp. S5-12, 2012.

[6] G. Healy, D. Dunstan, 1. Salmon, E. Cerin, 1. Shaw, P. Zimmet, and N. Owen, "Objectively measured light-intensity physical activity is independently associated with 2-h plasma glucose," Diabetes Care, vol. 1384-1389, 30.

[7] P. Freedson, H. Bowles, R. Troiano, and W. Haskell, "Assessment of physical activity using wearable monitors: recommendations for monitor calibration and use in the field.," Medicine and Science in Sports and Exercise, vol. 44(1 Suppl 1), pp. SI-4, 2012.

[8] P. Weyand, B. Smith, M. Puyau, and N. Butte, "The mass-specific energy cost of human walking is set by stature.," J Exp Bioi, vol. 213 (Pt 23), pp. 3972-9, 2010.

[9] H. Vathsangam, A. Emken, D. Spruijt-Metz, T. E. Schroeder, and G. S. Sukhatme, "Determining Energy Expenditure From Treadmill Walking Using Hip-Worn Inertial Sensors: An Experimental Study," IEEE Transactions on Biomedical Engineering, vol. 58, pp. 2804-2815, October 2011.

[IO] H. Vathsangam, B. Emken, E. Schroeder, D. Spruijt-Metz, and G. Sukhatme, "Hierarchical linear models for energy prediction using inertial sensors: a comparative study for treadmill walking," Journal of Ambient Intelligence and Humanized Computing, pp. 1-12, 2012. 1O.1007/s 12652-012-0150-y.

[11] RESL, "Movement Trackr Android Application https://github.comlmobilesensing-usc/MovementTrackr."

143