A Telemedicine Home Care Based Activity Monitor ...

0 downloads 0 Views 356KB Size Report
apposite for telemedicine home care settings. The proposed .... best position found by the particle so far, and having the highest scoring .... 1993; 25: 628-42. [3] Folsom AR, Jacobs DR Jr, Caspersen CJ, Gomez-Marin O, Knudsen J. ... [7] Ainsworth BE, Sternfeld B, Richardson MT, Jackson K. Evaluation of the Kaiser ...
A Telemedicine Home Care Based Activity Monitor Device D’Addio Gianni

Cimmino Pasquale, Carlo Manna, Arpaia Pasquale

Bioengineering Dept. of S. Maugeri Foundation, Rehabilitation Institute of Telese Terme, Italy [email protected]

Engineering Department of University of Sannio, Benevento, Italy [email protected], [email protected], [email protected]

Abstract — A minimally invasive low-cost actigraph, with simple on-line movements features extraction, but with performance enhanced by off-line post-processing, is proposed. Despite its cost significantly lower, the proposed actigraph has performance similar to commercially available solutions. Moreover, its open downloading system and wireless integration capabilities make it apposite for telemedicine home care settings. The proposed device has been characterized experimentally by comparing its performance with an emerging gold standard in clinical monitoring physical activity assessment. Keywords-component; formatting; style; styling; insert (key words)

I.

INTRODUCTION

Mobility is a basic physical requisite for an adequate quality of life, for both healthy and diseased people. Its assessment is a field of interest for Rehabilitation Medicine, a clinical discipline where the best possible restoration of both mobility and quality of life represents an important goal of many intervention programs. The International Classification of Functioning, Disability, and Health of the World Health Organization gives a central role to the identification and classification of the capability to have an adequate quality of life, monitoring and classifying physical activity (PA) abilities in performing daily functional motion tasks. Given the complex set of behaviors and heterogeneous nature of PA, many different types of instruments have been devised to measure mobility, mainly one based on the individual recording the amount of activity, and others based on instrumental monitoring. Individual recording techniques are based on both patient’s report (such as self- self-administered diaries, logs, and recall surveys) and recurrent exams by professionals or trained staff [1-7]. Qualitative information, gathered from these individual recording, has been converted often into estimates of Energy Expenditure (EE, measured in in kilo-calories, joules or metabolic equivalents METs), or some other summary measures, that can be used to categorize or rank people by their physical activity level. However, these approaches although of simple use, often exhibited inaccurate PA measurements [8,9]. Direct instrumental monitoring of PA under free-living conditions is usually performed by requiring the subject to wear specific devices like pedometers, actigraphs, or

movement recorders, mainly based on sensors’ technologies of one or more mono-multi-axial accelerometers. An accelerometer is a device producing an electrical output (i.e. charge, voltage, current or change of resistance) proportional to the acceleration, typically expressed in m/s2 or in g-values. Modern accelerometers are typically micromachined silicon sensors based on the detection of the displacement experienced by a small mass linked to a frame by beams when the sensor is subjected to an acceleration: the applied force, hence the acceleration, can be derived from the measure of the deflection. Piezo-resistive and variable capacitance accelerometers, very frequently used in human movement applications, respond to accelerations due to movement as well as to gravitational acceleration. The static response of these accelerometers reflects the orientation of the accelerometer with respect to gravity and can be used to compute the angle relative to the vertical of the sensor and, consequently, of the body segment on which it is located Commercially available pedometers are generally affected by limited sensitivity in detecting low-speed movements (for instance, while moving around the house), and cannot discern activities not involving locomotion [10-12]. On the other side, movement recorders devices, against the possibility of very accurate movement (walking, running, climbing and descending stairs, etc.) and posture (standing, sitting, lying) classification, suffer of the disadvantages of very high costs, complexity and wearing discomfort of the sensors sets, candidating these devices only to limited ambulatory monitoring [13-16]. Actigraphs seem the most indicated devices for practical long periods activity recordings, like in telemedicine applications. Although they lacks the ability to identify accurately the type of movements or postures, like the more complexes movement recorders, they are very easy to use and generally able to rank different PA levels [17], also if the EE derived by these devices has been often questioned [18-21]. Despite actigraphs’ costs are lower than those of movement recorders, commercially available actigraphs are still not cheaply available for home care settings, with closed proprietary off-line downloading system and generally without wireless integration capabilities in home LAN solutions. Aim of the paper is developing and testing a minimally invasive, low-cost actigraph with simple on-line movements features extraction, but with performance enhanced by off-line

post-processing, in a wireless home care LAN setting. At this purpose, the proposed device has been tested by comparing its performance with the BodyMedia SenseWear Armband, actually recognized as an emerging gold standard in clinical monitoring PA assessment [22-26]. II.

MATERIALS AND METHODS

A. The ACRHOME device The ACtivity Record at HOME unit (ACRHOME) has been realized by a tri-axial linear accelerometer component ST LIS3V02DQ ST detecting normal body motion accelerations up to 2 g. The device has an on-board ADC low power (0.5 mA), with a 40 Hz decimation factor and an high pass filter with cut-off frequency of 10 Hz to reduce output offset; it may be configured to generate wake-up/free-fall interrupt, with programmable acceleration threshold, to switch the system from sleeping mode when mechanically elicited. Since ACRHOME is one of the devices of a wireless domotic and home care multi-parametric monitoring system elsewhere described [27, 28], the unit is equipped by a Texas Instrument CC2480 multi-channel RF transceiver, designed for home LAN ZigBee network. A Texas Instruments MSP 430 microcontroller includes all the hardware and software capabilities required (i) to handle the accelerometer: I2C data interface, high frequency filtering movements (such as vibrations, out of the 0.25-2.5 Hz frequency band), data processing; (ii) develop an entire wireless communication protocol. The small dimension (less than 2x2cm) and light weight (120 g) device in located inside a plastic bracelet that can easily worn by the patient at wrist.

Fig.1 - Texas Instruments eZ430-RF2480 ZigBee module assembled with ST LIS3V02DQ accelerometer I2C interfaced with MSP430F2274 microcontroller unit

B. The ACRHOME pre-processing algorithm The device, considering a unique acceleration signal as the sum of the accelerations along the three axes, quantifies the following parameters, over each 1 minute time epoch [29]. Threshold Crossing (TC) is measured by recording a count each time the transducer signal crosses a defined threshold voltage regardless of whether the voltage is increasing or decreasing. Counts are then accumulated for each epoch and stored in the device’s memory.

Time above Threshold (TAT) is obtained by summing the time that the signal exceeds a previously defined acceleration threshold. At the end of each epoch, the value is stored in the device’s memory. Integrated Activity (IA) is computed by summing the deviations from 0,0 V (i.e. the absolute value of the voltage) during the epoch and storing the value at the end of the epoch. The matrix of all epochs times, TC, TAT and IA values is hourly wireless transmitted to a the local central unit of the home LAN ZigBee network for further off-line PA classification. If the device is out the network coverage because the subject wearing the ACRHOME device in temporally outdoor, the unit will regularly continue storing on board the epochs’ parameters, postponing data downloading when under home LAN coverage. C. The ACRHOME off-line PA classification algorithm One minute PA epochs, quantified by TC, TAT an IA parameters and stored on the local central unit of the home LAN ZigBee network, are remotely accessed by a GPRS/Internet connection from a medical control center where physicians can have an updated monitoring. Off-line post processing allows a PA epoch’s classification based on the benchmark with the Armband system. According to the Armband channels export data, each 1-minute PA epoch has been classified with one of the following labels of Table I. TABLE I. Label

PA classification

EPOCHS CLASSIFICATION LABELS Related EE in metabolic equivalents

N

Sleep

S

sedentary PA

3 < Mets

M

moderate PA

3 < Mets < 6

V

vigorous PA

6 < Mets < 9

W

very vigorous PA

9 < Mets

Epoch classification of ACRHOME data is obtained by a weighted linear combination of TC, TA and IA values by means of a Particle Swarm Optimization (PSO) inspired technique [30]. PSO is a modern artificial intelligence method derived as imitation of the birds flocking behavior and have been applied in a variety of data mining and optimization problems in continuous landscape like power systems or composite beam structures and operation research problems [31-33]. In particular, PSO is a population-based stochastic optimization technique inspired by the social behaviour of bird flocking. The system is initialized with random solutions ("particles") with an assigned random "velocity" vi, and then "flown" through the parameter space ui ∈ ℜ n . Each i-th particle (i.e. a potential solution) "flies” within its search space, conditioned by values of its personal best position pb (i.e., the best position found by the particle so far, and having the highest scoring according to a fitness function), and the global best position G (i.e. the best position found by all other) described by the authors [34], is based on an empirical knowledge preventively acquired from generic experiments

and then used to predict the condition of following observations. At this purpose the whole experimental data set of 10 recordings, described in the following paragraph, has been split in two subsets of 5 recordings, with a first set of learning data for PSO based parameters tuning and a second set over which epoch’s classification has been performed. III.

EXPERIMENTAL RESULTS

A. Experimental setting The system has been tested on a 10 healthy normal subject population (males, 35±8 years old). Subjects underwent a 24-hour activity monitor recordings contemporary wearing both ACRHOME bracelet and Armband systems respectively on the wrist and on the arm of their dominant side. After a time synchronization checking between the two internal clocks, the devices have been mounted between 8 and 9 A.M. of the first day and switched off at the same hour of the following day. The configuration parameters for the PSO were c1 = c2 = 0.5, to give equal weight to the mechanisms of exploration and exploitation of PSO, while the termination criteria was based on maximum number of iterations (max 2,000 runs). A reproducible, realistic and mixed PA was achieved by instructing the subjects to follow the general planned time scheduled in Table II. TABLE II. Period

Interval times

SCHEDULED DAY LOG Allowed Physical Activity

Related EE in metabolic equivalents

A1

from 9 to 12

moderate to very vigorous PA

≥ 3 Mets

S1

from 14 to 16

sedentary PA

< 3 Mets

A2

from 16 to 19

moderate to very vigorous PA

≥ 3 Mets

S2

from 21 to 23

sedentary PA

< 3 Mets

N

from 23 to 7

sleeping period

During the active time periods (A1, A2) subjects were allowed to spend time out of the house and required to perform two times a modified version of the protocol described by Bussmann [35] consisting for about 40 minutes of a sequence of activities/postures in this order: 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11.

Standing Sitting Lying Walking a corridor Walking ascending a flight of stairs Walking descending a flight of stairs Resting (standing) Walking on a treadmill at 4 Km/h Resting (standing) Walking on a treadmill at 6 Km/h Resting (sitting)

During the sedentary time periods (S1, S2) subjects were asked to maintain a PA level lower than 3 Mets (reading, listening music or watching TV, using telephone or PC), while remaining interval times between A,S,N-type phases has been considered transition times (T) with unclassified PA (dressing, undressing, preparing meals and eating, personal needs tec…). B. Experimental PA classification’s benchmarking In the following Table III, in the third and forth columns are shown the epochs’ time classification in percentage of the 24 hour elapsed time, reported as mean and standard deviation of the subset tested recordings, respectively by means of ACHROME and ARMBAND TABLE III. Day log period

PA CLASSIFICATION BENCHMARKING

Epoch type

ACRHOME Daily %

ARMBAND Daily %

Accuracy

p-value t-test

W

2,4±1,5

4,2±0,8

56,8

0,04*

V

9,2±1,8

8,3±2,0

89,2

0,48

M

13,4±3,1

12,5±3,5

92,8

0,68

S

S

14,8±4,3

16,7±3,9

88,9

0,49

N

N

35,2±8,7

33,3±9,1

94,5

0,74

S

10,4±1,6

8,3±1,2

75,4

0,04*

M

15,2±

16,7±

87,7

0,17

A

T

The benchmark evaluation between the two devices has been studied in terms of accuracy, defined for each epoch type as the percentage of agreement between the ACHROME versus the ARMBAND classification of each 1-minute epoch. ACRHOME results showed 1) an underestimation of very vigorous activity (W) corresponding to a related overestimation of vigorous (V) and moderate (M) epochs in the active periods; 2) an overestimation of the sleeping epochs (N) respect sedentary epochs (S) probably due to the fact that the ARMBAND is able to discriminate also sedentary lying but non sleeping time periods; 3) generally lower performance during transition periods, probably due to fast changing movements patterns switching between sedentary and moderate activity. Accuracy values mainly ranged between 87 and 93% with a highest value of 94.5% reached during the sleeping period, and the lowest one of 56.8% during the very vigorous activity. IV.

DISCUSSION

The developed ACRHOME device showed very minimally invasive impact on free-living daily activity of the subjects under test. They were simple able to wear the bracelet at wrist except than during personal hygiene caring times, with a continuous recording also during outdoor activities. The device clearly exhibited a lower sensitivity compared to the BodyMedia SenseWear Armband, especially during most critical conditions like very vigorous activity or fast changing movement patterns, unable to discriminate also between lying from sleeping times. This is mainly due to the fact that the BodyMedia SenseWear Armband, utilizing a proprietary multisensory array including in addition to a 2-axis

accelerometer, heat flux sensor, galvanic skin response sensor, skin temperature sensor, and a near-body ambient temperature sensor is not just a simple activity monitor, but a real metabolic holter. Anyway, although on the small sample size of normal subjects studied, in can be observed that epochs ACRHOME misclassifications lead to a not significantly different overall PA assessment. Particularly, performing a paired t test between the different daily epoch’s percentage classified by the two devices for each subject, their differences appeared statistically different (p