Validity and reliability of measurements obtained with an "activity monitor" in people with and without a trastibial amputation.

Key Words: Accelerometry, Ambulatory monitoring, Mobility, Physical activities, Validity.

Locomotion, ambulation, and mobility are important aspects of rehabilitation.[1-8] Many techniques are used for acquisition of mobility data, including the use of' questionnaires,[9-14] observation,[10,15-17] diaries,[18-19] kinetic and kinematic systems,[16,20,21] mechanical and electronic motion sensors,[17,22-25] and types of activity monitors.[26-32] The selection of a technique depends on, among other variables, the kind of information required. If unobtrusive, reliable, and valid measurements are required for a large and specific set of mobility activities during normal daily life in a person's own environment, current techniques fail to some extent. We therefore developed an "Activity Monitor" (AM),(*) an instrument that can be used for long-term monitoring of ambulatory activity by use of accelerometer signals and for assessment of the quantity (when, how long, how often) and quality (how performed) of several mobility activities. These activities include stationary activities (ie, standing, sitting, and 3 different modes of lying), movement-related activities (ie, walking, climbing stairs, cycling, and using a wheelchair), and the transitions between the stationary activities. Among other types of subjects, our studies will include people with amputation of the leg, because restricted mobility is a major problem tot these persons.

Until now, the development of the instrument comprised the selection of the type, number, and location of sensors and the optimization of analysis algorithms,[33,34] based on data for subjects without disease or impairment. The validity of measurements obtained with the instrument, however, needs much attention (ie, whether AM-derived measurements actually reflect the subject's activities). Validity may be influenced by factors such as age, sex, height, weight, disease or impairment, phase of rehabilitation, amputation level, and setting. The aim of this study, therefore, was to investigate the reliability and validity of AM-derived measurements obtained for persons with and without an amputation. The main research question was: Can the type and duration of activities and the number of transitions be validly measured by the AM? A secondary research question was: Does the AM function at the same level of accuracy (1) when measurements are repeated, (2) when the instrument is used with different subjects, and (3) when the instrument is used with persons with and without a transtibial amputation?

Method

Activity Monitor

The AM consists of accelerometers, a portable data recorder, and a computer with analysis programs. In this study, IC-3031 uniaxial piezo-resistive accelerometers([dagger]) (1.5 x 2 x 0.5 cm) were used. The signals of these sensors consist of both a component of the gravitational acceleration and a component of other accelerations, it present.[33,34] The magnitude of these components depends on the direction of these accelerations with regard to the sensitive axis of the sensor and their magnitudes.

Four sensors were fixed on the skin by means of double-sided tape. Two sensors were attached on the front of the thighs, halfway between the anterior superior iliac spine and the upper side of the patella, and 2 sensors were attached on the lower part of the sternum, perpendicular to each other (Fig. 1). The trunk sensors were also held in place by means of a rubber belt. All accelerometers were attached as parallel as possible to the vertical or horizontal plane; a maximum deviation of 15 degrees was allowed. Each accelerometer was connected to a portable Vitaport 1 data recorder([dagger]) (13 x 9 x 4 cm, 480 g, battery included) by a cable (under the clothes) and a Lemo-jacket.([dagger]) The recorder was worn on a belt around the subject's waist. Power was delivered by a rechargeable battery (270 mAh, 4.8 V). Raw signals were digitally stored on a removable memory card, with a sampling frequency of 25 Hz.

[Figure 1 ILLUSTRATION OMITTED]

After the measurements, the data were downloaded to a Macintosh IIci computer([double dagger]) for analysis. Although the signals of 4 sensors were measured, the signal of the left or amputated leg was not used in the analysis. The signal of the other leg was measured to study the quality of walking. The data were analyzed by means of Vitagraph([sections]) Signal Processing and Inferencing Language (SPIL).[35],([parallel])

The output of the AM is the automatic 1-second selection of one type of activity (Fig. 2). To achieve this output, 2 types of signals were derived from each sensor signal: (1) a low-pass filtered (0.5 Hz) signal (3 HPRS signals) and (2) a successively high-pass filtered (0.5 Hz), rectified, arid smoothed signal (3 HPRS signals), The LP signals were used to distinguish 5 stationary activities, because these activities have a unique combination of 3 LP signals. The HPRS signal of the thigh was used to distinguish between movement-related and stationary activities. Movement-related activities are characterized by variability of the accelerometer signal. The more "energetic" an activity, the more variable the accelerometer signal and the higher the value of the HPRS signal. The way in which the movement-related activities can be distinguished from each other is still under investigation. This study, therefore, was restricted to the global categories "stationary" and "movement-related," the 5 stationary activities, and the transitions.

[Figure 2 ILLUSTRATION OMITTED]

Reference Method

Videotape recordings were chosen as the reference method, or standard. During all measurements, videotape recordings (with a video clock) were made, together with the monitoring of AM output. To allow a correct comparison of Ore videotape and AM data, the timing of both instruments was synchronized. The videotape recordings were made and analyzed by the same person (a medical student during her research traineeship), independent from the AM analysis. In a later study,[36] we investigated the interrater reliability of data from the videotape analysis. An overall agreement of 99.7% was found between 2 raters, indicating the reliability of the data from the videotape analysis.

The classification categories of the videotape analysis were the same as the classification categories of the AM analysis, and the output signals of both instruments had the same 1-second time resolution. The guidelines for the videotape analysis, however, were different horn the guidelines for the AM analysis. The videotape analysis of lying, sitting, and standing was based on the presence and position of supporting surfaces, whereas the detection of posture with the AM was based on the angular position of the thighs and the trunk. Furthermore, in the videotape analysis, only cyclic activities (walking, climbing stairs, cycling) were determined as movement-related, whereas the AM may also determine noncyclic activities as movement-related. After synchronization to the signals in the AM file, the videotape recording time was converted to sample numbers. These sample numbers and their corresponding category codes were edited so that they could be transferred to a signal in the AM file (Fig. 2).

Protocol

The measurements were taken in a seminatural setting in the occupational therapy department, in which a complete (representative) apartment had been installed. During the measurements, the subjects performed several functional activities, including dressing, going to bed, preparing breakfast, peeling and cooking potatoes, watching television, reading a newspaper, shopping on another floor after climbing a stairway, and riding a bicycle (using a wheelchair was not included). These activities were selected by an occupational therapist. Bet0re the measurements, the protocol was explained to the subjects. When the measurements were taken, subjects were allowed to do the activities in their own way and at their own pace. The measurements were planned to last approximately 45 minutes.

Subjects

The following inclusion criteria were used for the subjects with an amputation: one-sided transtibial amputation, recent ([is less than] 6 months) discharge from an outpatient rehabilitation clinic, age greater than 18 years, no use of assistive devices, ability to complete the protocol, and no diseases or impairments disturbing locomotion. A rehabilitation specialist selected 4 men from a file of discharged patients (mean age = 32 years, SD = 17.7, range = 19-57; mean height = 1.82 m, SD = 0.05, range = 1.75-1.85; mean weight = 74 kg, SD = 9.2 range = 63-83). For each patient, a person without an amputation and of the same sex, age ([+ or -] 10 years), weight ([+ or -] 10 kg), and height ([+ or -] 0.10 m) was selected. The patients performed the protocol once, and the comparison subjects performed the protocol twice on different days to determine test-retest reliability. A total of 12 measurements were performed.

Data Analysis

Both the AM output signal (with AM activity category codes) and the videotape analysis signal (with the videotape activity category codes) had a time resolution of 1 second. Every second, the codes of both signals could be compared. In this way the number of corresponding and noncorresponding counts (1 count = 1 second) and the agreement scores could be calculated. Because the videotape analysis can be regarded as a standard, the following agreement scores--as validity measures of the AM--were used (research question 1):

1. Agreement: the percentage of agreement between all samples of videotape recording and AM data. Agreement was calculated according to the equation: agreement = (number of identical samples of videotape recording data and AM data/total number of samples) x 100%.

2. Sensitivity: the degree to which each videotape activity category (representing the activities actually performed) was detected correctly by the AM. Sensitivity was calculated according to the equation: sensitivity for videotape activity category A = (number of identical samples of videotape recording data and AM data when videotape activity category is A/total number of samples for videotape activity category A) x 100%.

3. Predictive value: the degree to which each AM activity category agreed with the videotape activity category (representing the activities actually performed). Predictive value was calculated according to the equation: predictive value of AM activity category A = (number of identical samples of videotape recording data and AM data when AM activity category is A/total number of samples for AM activity category A) x 100%.

The 1-second output of the AM and videotape recording analysis allowed calculation of duration (in seconds) per activity. The number of transitions within each transition category was calculated from identified changes in posture. All calculations and comparisons were done automatically by means of SPIL software.

Simple descriptive statistical measures, such as weighted (corrected for duration of activities) mean and standard deviation, were used to describe group results. The Wilcoxon matched-pairs signed-rank test was used to show systematic differences in results between the videotape recording analysis and the AM analysis. The Mann-Whitney U test was used to show systematic differences in results between the patient group and the first measurement of the comparison group. All statistical analyses were done with SPSS 5.0 for MS Windows.(#) A probability value of P [is less than] .05 was considered to indicate a significant effect.

Results

The overall agreement between the videotape data and the AM output was 90%. The overall (weighted) mean of most sensitivities and predictive values equaled or exceeded 90% (Tabs. 1 and 2). The overall sensitivity for lying on the side and for movement-related activities was somewhat lower (88% and 85%, respectively), as was the overall predictive value of the AM activity categories of sitting and movement-related activities (88% and 89%, respectively). During 890 seconds, standing (determined by videotape analysis) was detected by the AM as a movement-related activity. During 916 seconds, however, movement-related activities (determined by videotape analysis) were detected by the AM as standing. Table 1 also provides insight on the distribution of activities during the measurements. No differences in distribution of activities existed between the groups, although the measurements in the patient group lasted, on average, longer than those in the comparison group (41 and 34 minutes, respectively).

Table 1. Number of Corresponding (Underlined) and Noncorresponding (Plain) Counts (1 Count = 1 Second) of Videotape Act (AM) Output Added for All Measurements(a)

                                          AM Output

                               Lying    Lying
                               on       on      Lying
Videotape Activity             Back     Side    Prone   Standing

Lying on back                  853          0       0          0
Lying on side                    0      1,057       0          5
Lying prone                      0          0      67          0
Standing                         0          0       0      8,755
Sitting                          0          0       0          0
Movement-related activity       27         20       5        916
Total                          880      1,077      72      9,676
Overall predictive value (%)    97         98      93         90

                                    AM Output

                                          Movement-
                                          Related
Videotape Activity              Sitting   Activity    Total

Lying on back                        53      15          921
Lying on side                       118      20        1,200
Lying prone                           0       1           68
Standing                             83     890        9,728
Sitting                           4,809     108        4,917
Movement-related activity           421   8,068        9,457
Total                             5,484   9,102       26,291
Overall predictive value (%)         88      89

                                Overall
                                Sensitivity
Videotape Activity              (%)

Lying on back                   93
Lying on side                   88
Lying prone                     99
Standing                        90
Sitting                         98
Movement-related activity       85
Total
Overall predictive value (%)    90

(a) The last column shows the overall sensitivity of the AM for each videotape activity category; the bottom row shows the overall predictive value of each AM activity category.

Table 2. Percentages per Measurement, Representing the Sensitivity (S) and Predictive Value (PV)(a)

                   Sensitivity and Predictive Value (%)

              Lying on     Lying on     Lying
Subject/      Back         Side         Prone        Standing
Measurement    S     PV    S      PV    S     PV     S     PV

H1M1                        99     99                 90     90
H2M1          100     97   100     98                 77     91
H3M1           94     91   100     91                 91     97
H4M1                        93     99                 77     90
               97     93    97     98                 85     92
H1M2          100     99    95    100                 92     91
H2M2          100     95    16     88                 77     96
H3M2           97     98                              90     92
H4M2                        99    100                 85     88
               98     98    70     99                 86     92
A1M1                        98     98   99    93      97     89
A2M1           61    100                              97     92
A3M1          1O0     97    95     95                 95     94
A4M1           99     96   100     98                 96     80
               86     97    99     98   99    93      96     89
(overall)      93     97    88     98   99    93      90     90

               Sensitivity and
               Predictive Value (%)

                            Movement-
                            Related
Subject/       Sitting      Activity     Agreement
Measurement    S     PV     S     PV     (%)

H1M1            97     95   86     87    91
H2M1            98     97   93     85    90
H3M1            95     82   91     90    92
H4M1            97     87   88     80    86
                97     92   90     85    90
H1M2            99     94   88     92    93
H2M2            98     74   94     86    85
H3M2            98     92   89     90    92
H4M2            94     91   90     88    90
                97     87   91     89    90
A1M1            98     97   80     93    92
A2M1            97     69   73     92    88
A3M1           1O0     87   80     92    92
A4M1           100     95   71     94    88
                99     86   76     93    90
(overall)       98     88   85     89    90

(a) Ellipsis indicates that the activity was not performed or detected. The weighted means for each subgroup are calculated. In the last column, the agreement per measurement is shown; in the last row, the weighted overall means are shown. Measurement code: H1 - H4 = subjects without amputation, A1 - A4 = subjects with transtibial amputation, M1 and M2 = measurements 1 and 2.

The AM slightly overestimated the total number of transitions compared with the videotape recordings (overall difference: +16 [+7%], P [is less than] .05, Tab. 3). The overall duration of activities determined by the AM did not deviate from the duration of activities determined by videotape analysis for standing (Tab. 4). The duration of sitting was overestimated by the AM (mean difference = 2.2%, P [is less than] .01).

Table 3. The Number of 6 Transition Types, Determined by Activity Monitor (AM) and Videotape Analysis (V)(a)

                       No. of Transitions

                  Lying-      Lying-       Sitting-
Subject/          Sitting     Standing     Lying
Measurement       V     AM    V      AM    V      AM

H1M1              0      0    1       1    0       0
H2M1              1      2    0       0    0       1
H3M1              1      1    0       0    0       1
H4M1              0      0    1       2    1       1
Total             2      3    2       3    1       3

H1M2              0      0    1       1    0       0
H2M2              1      1    0       0    0       0
H3M2              1      1    1       1    1       1
H4M2              0      0    1       1    1       1
Total             2      2    3       3    2       2

A1M1              1      1    0       0    1       1
A2M1              2      2    0       0    2       2
A3M1              1      1    0       0    1       1
A4M1              0      0    1       1    0       1
Total             4      4    1       1    4       5

Total (overall)   8      9    6       7    7      10

                       No. of Transitions

                   Sitting-    Standing-    Standing-
Subject/           Standing    Lying        Sitting
Measurement        V      AM   V      AM    V        AM

H1M1               8      9   1        2     9        9
H2M1               9      9   1        1     8        8
H3M1               9      9   1        0     8        9
H4M1               7      8   0        1     8        9
Total             33     35   3        4    33       35

H1M2               7      7   1        1     7        7
H2M2              12     12   1        1    11       11
H3M2               8      8   1        1     8        8
H4M2               5      6   0        0     6        7
Total             32     33   3        3    32       33

A1M1               8      7   0        0     8        7
A2M1               8     10   0        0     8       10
A3M1               5      5   0        0     5        5
A4M1               7      8   1        0     7        9
Total             28     30   1        0    28       31

Total (overall)   93     98   7        7    93       99

                  No. of Transitions

                  Total
Subject/          V      AM
Measurement
                  19     21
H1M1              19     21
H2M1              19     20
H3M1              17     21
H4M1              74     83
Total             16     16

H1M2              25     25
H2M2              20     20
H3M2              13     15
H4M2              74     76
Total             18     16

A1M1              20     24
A2M1              12     12
A3M1              16     19
A4M1              66     71
Total            214    230

Total (overall)

(a) The data are shown per measurement, for each subgroup, and for all measurements together. Measurement code: H1 - H4 = subjects without amputation, A1 - A4 = subjects with transtibial amputation, M1 and M2 = measurements 1 and 2.

Table 4. Duration (as Percentage of the Measurement Time) of Each Activity Category, Determined by Activity Monitor (AM) and Videotape Analysis (V)(a)

                          Duration (%)

                 Lying        Lying         Lying
Subject/         on Back      on Side       Prone
Measurement      V      AM     V      AM     V     AM

H1M1             0.0    0.0   9.0     8.9   0.0   0.0
H2M1             1.7    1.8   5.7     5.9   0.0   0.0
H3M1             2.6    2.7   2.5     2.8   0.0   0.0
H4M1             0.0    0.0   9.3     8.8   0.0   0.0
X                1.0    1.1   6.7     6.6   0.0   0.0

H1M2             5.3    5.4   4.3     4.1   0.0   0.0
H2M2             3.9    4.1   6.7     1.2   0.0   0.0
H3M2            10.9   10.8   0.0     0.0   0.0   0.0
H4M2             0.0    0.0   9.3     9.2   0.0   0.0
X                5.1    5.1   5.0     3.5   0.0   0.0

A1M1             0.0    0.0   2.7     2.7   3.0   3.1
A2M1             4.9    3.0   0.0     0.0   0.0   0.0
A3M1             7.3    7.6   1.9     1.9   0.0   0.0
A4M1             4.6    4.8   5.8     5.9   0.0   0.0
X                4.2    3.8   2.4     2.5   0.7   0.7

X (overall)      3.5    3.3   4.6     4.1   0.3   0.3

                          Duration (%)

                                             Movement-
                                             Related
Subject/        Standing      Sitting        Activity
Measurement     V      AM     V      AM      V      AM     Total

H1M1            36.9   36.8   23.0   23.4    31.1   30.9   100
H2M1            29.9   25.2   19.0   19.1    43.7   48.1   100
H3M1            44.5   41.9   10.9   12.6    39.6   40.1   100
H4M1            35.3   29.9   17.5   19.6    37.8   41.7   100
X               36.6   33.5   17.8   18.8    37.9   40.0   100

H1M2            30.2   30.7   26.4   27.6    33.8   32.3   100
H2M2            27.4   22.2   21.1   27.6    40.9   44.9   100
H3M2            33.5   33.0   15.5   16.6    40.1   39.6   100
H4M2            34.3   32.9   15.1   15.7    41.3   42.2   100
X               31.3   29.6   19.6   22.1    39.0   39.8   100

A1M1            39.4   43.0   25.5   26.0    29.4   25.1   100
A2M1            49.2   51.6   14.4   20.3    31.6   25.2   100
A3M1            36.9   37.2   23.5   26.9    30.4   26.4   100
A4M1            40.3   48.1   13.0   13.7    36.3   27.5   100
X               42.0   45.5   18.7   21.6    31.9   26.0   100

X (overall)     37.0   36.8   18.7   20.9    36.0   34.6   100

(a) The data are shown per measurement, and weighted means are calculated for the subgroups and for all measurements together. Measurement codes: H1 - H4 = subjects without amputation, A1 - A4 = subjects with transtibial amputation, M1 and M2 = measurements 1 and 2.

The standard deviation of the difference in agreement between the first and second measurements for the comparison subjects was 3.9% (range = -5%-4%). The standard deviations for the percentage of agreement ranged from 2.3% (range = 88%-92%) in the patient group to 3.6% (range = 85%-93%) for the second measurement in the comparison group. When all measurements were included, the standard deviation was 2.6% (range = 85%-93%).

To examine differences in functioning of the AM in persons with and without transtibial amputation, variables were compared between the patient group and the comparison group (first measurement). The agreement did not differ between groups. Some percentages, however, for sensitivity and predictive value differed (Tab. 2). The sensitivity for standing and the predictive value for standing were higher in the patient group than in the comparison group, and the sensitivity for movement-related activities was higher in the comparison group than in the patient group (P [is less than] .03).

The functioning of the AM, expressed as the correct determination of the number of transitions, did not differ between groups (Tab. 3). The duration of movement-related activities was overestimated by the AM in the comparison group (+2.1%) and underestimated in the patient group (-5.9%). The duration of standing was underestimated in the comparison group (-3.1%) and overestimated in the patient group (+3.5%).

Discussion

The agreement scores in this study were generally within a range of error of 0% to 10%. A comparison of the results of our study with those of other studies and instruments is not possible because the different instruments used in these studies discriminate among different activity categories,[26-28,37] the protocols consisted of different activities,[29,32] or validity was calculated following another or unknown method.[30,31,38] Kiani et al[38] have developed the "Ambulatory Monitoring of Motor Activities" (AMMA) system using accelerometers and an artificial neural network. The set of activities that can be detected with the AMMA system are similar to the activity categories of the AM. Although the technique seems promising and 95% reliability is reported, the validation technique used is questionable. The measured accelerometer signals were the input of both the AMMA system and the reference method (visual interpretation of the signals). Walker et al[37] reported a validation study of an activity monitor based on mercury switches and accelerometers. Validation was studied in terms of steps counted; validation of body positions was not reported. Stock and colleagues[30,31] used a "microcomputer-based system for the assessment of postoperative fatigue" consisting of a posture timing module, an activity module, and a heart rate module. Although they reported maximum error percentages of about 5%, they did not clearly describe how these percentages were obtained, and the results, therefore, are difficult to interpret.

Anastasiades and Johnston[26] used electromyography to discriminate between stationary and movement-related activities. Stationary activities could not be distinguished from each other, and interindividual comparisons appeared to be difficult. Fahrenberg et al[29] applied accelerometer sensors and a hydrostatic tube to monitor their subjects' ambulatory activity. The same sensors were applied by Tuomisto et al,[32[ who also used electromyography. In both studies, only a small number standardized activities were performed to determine validity.

Instruments for measuring walking periods were validated by Bassey et al.[27] Although the measurements they obtained showed reasonable validity, the instruments they used measured only walking and stair-climbing performance; these instruments were not designed to measure cycling and different body positions. Several workers[17,39-41] used accelerometry to distinguish between stationary and movement-related activities and to determine the level of activity. Due to the limited scope of these instruments, comparison with the AM is not useful. The instrument used by Diggory et al[28] is designed with a tilt switch to measure time in the upright position. Although their validity study gave good results, possible errors were reported for measurements obtained when the subjects were lying prone, cycling, and climbing stairs. Furthermore, movement-related activities formed no part of the output of the instrument.

Some arguments against the present study can also be posed. The activities were prescribed and limited in number, and they were performed in the same environment--an artificial environment, not the subjects' home environment. The subjects selected for the patient group had transtibial amputations and had finished rehabilitation. Furthermore, they were relatively good walkers, and 3 of the 4 subjects in this group were fairly young. In our view, the generalizability of the results of our study to unsupervised measurements is enhanced by some other characteristics of the study. The measurements were done in a seminatural setting, and the activities were functional and selected by an occupational therapist who did not participate in the study. The subjects could perform the activities in their own manner, and the method for assessing the validity of the measurements was critical. If the AM is to be used in the real-world environment of patients (eg, with external vibrations due to car or train, extension of the kinds of movements and postures[17,42]) and with patients with other movement patterns[43] (eg, other impairments, other amputation levels, other phases of rehabilitation, different ages), however, further study of the reliability and validity of measurements obtained with the device is needed. In later validation studies of the AM, used with patients with failed back surgery in their own environment[44] and with subjects without known disease or impairment in the setting of a psychophysiological study,[36] results similar to those of our study were obtained. Overall percentages of agreement of 87% and 88% were found, respectively, supporting the validity of the measurements and the robustness of the AM.

While we were taking the measurements in our study, there were no problems with either the sensors or the recorder system, and neither the cables nor the recorder interfered with any of the subjects' activities. We observed, however, a tendency of the sensors (especially the trunk sensors) to come somewhat loose from the skin, probably due to chest hair, perspiration, or the rubber belt. In later studies,[36,44] therefore, we taped the sensors onto the skin with other material (eg, Kushionflex(**)). The method of fixation of the sensors is still a focus of our research interest.

Recently, we obtained measurements (1-2 days) with a recorder of slightly greater size and weight than that used in the present study and found that the measurement system may cause some discomfort. The system cannot be used in a wet environment (eg, while bathing or taking a shower). Some people find the recorder or the cables disturbing while sleeping or while dressing or undressing, and some people dislike being seen wearing the instrument. Therefore, we are investigating methods that will allow patients to easily attach and remove the recorder and sensors themselves, without compromising the measurements. Reducing the weight and size of the recorder may also increase comfort and applicability. Five prototypes of a new recorder (RAM(([dagger]) ([dagger])), of approximately the same size and weight as the recorder used in our study, have been developed to enable continuous measurement of accelerometer signals for more than 48 hours (compared with approximately 2 hours for the recorder used in our study), without replacement of flash cards or batteries. The recorder receives power from 4 penlight batteries and allows measurement of up to 8 signals. Data are stored on a flash card of 40 MB (or more). Reading the data from the flash card takes about 1 minute. The analysis of data (for a 2-day measurement) takes less than 1 hour. Generally, we start the measurements in the patient's home, and the data are downloaded in our laboratory. Downloading the data by telephone is now an option.

The AM was not free of errors. The investigation of even small errors can be used to increase knowledge about the functioning of the AM. Relatively often, standing (according to videotape analysis) is detected as a movement-related activity by the AM, whereas some movement-related activities (according to videotape analysis) are detected as sitting or standing by the AM (Tab. 1). The distinction between the 2 global categories of stationary and movement-related activities occurs early in the analysis program. Errors at this point have irreversible consequences for subsequent phases of the activity detection. Furthermore, small time shifts between videotape and AM signals, infrequent activities that were difficult to analyze using videotape recordings, and timing inaccuracies in videotape analysis (as shown in Fig. 2) may have some effect on these errors. The interrater reliability of data from the videotape analysis in a similar study performed later,[36] however, showed a percentage of agreement of 99.7%. The effect of videotape analysis errors on the data will be small, and that result showed that the videotape recording can be used as a standard. Other causes were shown to be more important.

The detection of an activity as stationary or movement-related strongly depends on an adjustable threshold applied on the HPRS signal of the leg, and the setting of this threshold is a matter of optimization. The detection of standing as a movement-related activity (eg, due to leg movement during standing, but not walking) and the detection of movement-related activities as standing (eg, due to shuffling) are quantitatively important misinterpretations. They lead to a decrease in overall agreement of about 6%. Generally, however, the threshold seemed well-chosen, as can be concluded from the small difference (-1.4%) between the duration of movement-related activities determined by AM and videotape recordings and the almost equal number of counts of the detection of movement-related activities as standing and of the detection of standing as a movement-related activity.

A movement-related activity was rather frequently (421 seconds) determined as sitting (Tab. 1). Cycling appeared to be the main activity that caused this discrepancy. During cycling, there were periods when the subjects did not move their legs. These periods were periods of movement-related activity according to videotape analysis, whereas the AM detected sitting. Furthermore, if the legs were moving during cycling, the acceleration energy was sometimes too low to cause the detection of movement-related activities. To quantify the effect of these errors, we studied the occurrence of these errors. The determination of a movement-related activity such as sitting was, for about 300 seconds of output, explained by these errors. If these errors had not occurred, the sensitivity of movement-related activities would have increased from 85% to 88% and the predictive value for sitting would have increased from 88% to 93%.

If decision errors for stationary and movement-related activities are excluded (ie, the "movement-related activity" row and column in Tab. 1 are not included in the analysis), the quality of the stationary activity detection can be assessed. The overall agreement then increases to 98%, and most sensitivities and predictive values equal, or come close to, 100%. Only the overall sensitivity for lying on the back (94%) and lying on the side (90%) and the overall predictive value of sitting (95%) are then clearly lower than 100%. This finding is due to 2 misinterpretations. In one comparison subject (H2M2, 118 seconds), lying on the side with the upper side of the trunk elevated and the legs rotated (while reading a book) was detected as sitting. In one patient (A2M1, 53 seconds), lying on the back (according to the criteria of the videotape analysis) with the trunk supported by a few pillows was also detected as sitting.

The total number of transitions was slightly, but systematically, overestimated by the AM. This finding was mainly due to overestimation of the sitting-to-standing and standing-to-sitting transitions (eg, foot on a chair while donning and doffing shoes and socks [videotape: standing; AM: standing-sitting-standing])and to overestimation of more complex transitions (eg, from standing to lying on the back via sitting and lying on the side). The duration of sitting also was overestimated. When data are corrected for the 2-fold misinterpretation of stationary activities and the determination of cycling as sitting, a considerable improvement is reached. Existing differences in duration (last row, Tab. 4) will almost disappear.

The differences in the agreement scores between the first and second measurements of the 4 comparison subjects were small, as were the differences in agreement scores among subjects within the same group, even though considerable individual differences in performance were observed. The mean agreement was the same for the patient group and the comparison group (90%). Comparison of the 2 groups, however, revealed some differences in results. In general (also after correction for the determination of cycling as sitting), movement-related activities were overestimated in the comparison group (+2.1%) but underestimated in the patient group (-5.9%). Standing, however, was overestimated in patient group (+ 3.5%) but underestimated in the comparison group (-3.1%). These differences between groups were statistically significant. The data suggest that the subjects with transtibial amputations in this study walked, cycled, and climbed stairs less energetically (ie, with less acceleration variation) than did the comparison subjects.

The potential of the AM can be increased by simultaneous measurement of, for example, electrocardiographic activity in quantifying strain, or markers. Furthermore, we assume that the accelerometer signals contain much information on the quality of the activities (how performed). Our research focuses especially on the quality of walking (spatiotemporal variables, stability).

There may be a considerable interday and intersubject variability,[14,27,45] which may differ by patient group. This variability will determine the number of days over which measurements should be taken and the number of subjects included in group studies. Future studies should investigate these 2 types of variability.

The AM is an instrument that provides data on the activities a patient actually performs during daily life. Generally, in many medical disciplines and also in physical therapy, the decisions about treatment and the evaluation of that treatment are increasingly attuned to and related to functional performance. For example, the effectiveness of physical therapy is determined not only by the change in joint mobility but also by changes in the performance of activities of daily living. The usefulness of the AM in physical therapy intervention, therefore, will depend on the treatment goals. The AM can be a powerful instrument in evaluative studies, especially when the formulated goals are related to the quantity of movement and postures. We believe that the AM will first be used in research. After making the instrument more dedicated, improving its usability, and reducing its costs, however, we expect that the AM will also be used in the practice of physical therapists.

Conclusion

Activity monitoring by means of accelerometry proves to be a promising method to obtain reliable and valid measurements of the activities a patient actually performs during daily life, which is essential in rehabilitation and physical therapy. Research with less obtrusive devices in real-world settings is now needed.

(*) The development started in 1992 as a joint project of Erasmus University Rotterdam (Department of Rehabilitation), University of Twente (Department of Electrical Engineering), and an industrial partner, all from the Netherlands.

([dagger]) Supplied by Becker Ingenieurburo, Karl-Seckinger, Strasse 48, D-7500 Karlsruhe 41, Germany.

([double dagger]) Apple Computer BV, Handelsweg 2, 3707 NN Zeist, the Netherlands.

([sections]) G Mutt. Department of Psychophysiology, University of Cologne, Cologne, Germany.

([parallel]) G Mutz, Department of Psychophysiology, University of Cologne. Cologne. Germany, and Will Martens, Phyvision, Kromstraat 3, Gemert, the Netherlands.

(#) SPSS Benelux BV, PO Box 115, 2200 AC Gorinchem, the Netherlands.

(**) Smith and Nephew Nederland, PO Box 535, 2130 AM Hoofddorp, the Netherlands.

([dagger]) ([dagger]) TEMEL, Instruments BV, PO Box 3011, 6460 HA Kerkrade, the Netherlands.

References

[1] van Bennekom CAM, Jelles F, Lankhorst GJ. Rehabilitation Activities Profile: the ICIDH as a framework for a problem-oriented assessment method in rehabilitation medicine. Disabil Rehabil. 1995;17:169-175.

[2] Brodzka WK, Thornhill HL, Zarapkar SE, et al. Long-term function of persons with atherosclerotic bilateral below-knee amputation living in the inner city. Arch Phys Med Rehabil. 1990;71:895-900.

[3] Collen FM, Wade DT, Robb GF, Bradshaw CM. The Rivermead Mobility Index: a further development of the Rivermead Motor Assessment. Int Disabil Stud. 1991; 13:50-54.

[4] Halstead LS. Activity monitoring as a clinical tool in rehabilitation medicine. In: Stott FD, Raftery EB, Sleight P, Goulding L, eds. Proceedings of the Second International Symposium of Ambulatory Monitoring, 1977. London, England: Academic Press Inc (London) Ltd; 1978: 247-258.

[5] Rozendal RH. Clinical gait analysis: problems and solutions? Human Movement Science. 1991;10:555-564.

[6] Torres M, Esquenazi A. Bilateral lower limb amputee rehabilitation: a retrospective review. West J Med. 1991;154:583-586.

[7] Wade DT. Measurement in Neurological Rehabilitation. Oxford, England: Oxford University Press; 1992.

[8] International Classification of Impairments, Disabilities, and Handicaps. Geneva, Switzerland: World Health Organization; 1981.

[9] Ayen TG, Montoye HJ. Estimation of energy expenditure with a simulated three-dimensional accelerometer. Journal of Ambulatory Monitoring. 1988;1:293-301.

[10] Bouchard TJ. Field research methods: interviewing, questionnaires, participant observation, systematic observation, unobstrusive measures. In: Dunette MD, ed. Handbook of Industrial and Organizational Psychology. Chicago, Ill: Rand McNally; 1976:364- 402.

[11] Burdorf A, Laan J. Comparison of methods for the assessment of postural load on the back. Scand J Work Environ Health. 1991;17: 425-429.

[12] Rakowksi W, Mor V. The association of physical activity with mortality among older adults in the Longitudinal Study of Aging (1984-19885. J Gerontol. 1992;47:M122-M129.

[13] Staples D. Questionnaires. Clinical Rehabilitation. 1991;5:259-264.

[14] Washburn RA, Montoye HJ. The assessment of physical activity by questionnaire. Am J Epidemiol. 1986;123:563-576.

[15] Carr EK. Observational methods in rehabilitation research. Clinical Rehabilitation. 1991;5:89-94.

[16] Geurts ACH, Mulder TH, Rijken RAJ, Nienhuis B. From the analysis of movements to the analysis of skills: bridging the gap between laboratory and clinic. Journal of Rehabilitation Sciences. 1991;4:9-12.

[17] Patterson SM, Krantz DS, Montgomery LC, et al. Automated physical activity monitoring: validation and comparison with physiological and self-report measures. Psychophysiology. 1993;30:296-305.

[18] Montgomery GK, Reynolds NC Jr. Compliance, reliability, and validity of self-monitoring for physical disturbances of Parkinson's disease: the Parkinson's Symptom Diary. J Nerv Ment Dis. 1990;178:636-641.

[19] Slender M, Doting A, Hense HW, et al. Veigleich zweier Methoden zur Erhebung der koerperlichen Aktivitaet. Soz Praventivmed. 1991:36: 176-183.

[20] Bowker P, Messenger N. The measurement of gait. Clinical Rehabilitation. 1988;2:889-897.

[21] Evans AL, Duncan G, Gilchrist W. Recording accelerations in body movements. Med Biol Eng Comput. 1991;29:102-104.

[22] Hunter GR, Montoye HJ, Webster JG, Demment R. The validity of a portable accelerometer for estimating energy expenditure in bicycle riding. J Sports Med Phys Fitness. 1989;29:218-222.

[23] Klesges RC, Klesges LM, Swenson AM, Pheley AM. A validation of two motion sensors in the prediction of child and adult physical activity. Am J Epidemiol. 1985;122:400-410.

[24] Klesges RC, Haddock CK, Eck LH. A multimethod approach to the measurement of childhood physical activity and its relationship to blood pressure and body weight. J Pediatr. 1990;116:888-893.

[25] Washburn RA, Cook TC, LaPorte RE. The objective assessment of physical activity in an occupationally active group. J Sports Med Phys Fitness. 1989;29:279-284.

[26] Anastasiades P, Johnston DW. A simple activity measure for use with ambulatory subjects. Psychophysiology. 1990;27:87-93.

[27] Bassey EJ, Fentem PH, Fitton DL, et al. Analysis of customary physical activity using long-term monitoring of ECG, and footfall. In: Stott FD, Raftery EB, Sleight P, Goulding L, eds. Proceedings of the Second International Symposium of Ambulatory Monitoring, 1977. London England: Academic Press Inc (London) Ltd; 1978:207-217.

[28] Diggory P, Gorman M, Helme R. An automatic device to measure time spent upright. Clinical Rehabilitation. 1994;8:353-357.

[29] Fahrenberg J, Mueller W, Foerster F, Smeja M. A multi-channel investigation of physical activity. J Psychophysiol. 1996;10:209-217.

[30] Stock SE, Johnston IDA. A system for monitoring activity at home in post-operative patients. J Ambul Mon. 1988;1:153-162.

[31] Stock SE, Clague MB, Johnston IDA. Post-operative fatigue: a real phenomenon attributable to the metabolic effects of surgery on body nutritional stores. Clin Nutr. 1991;10:151-157.

[32] Tuomisto MT, Johnston DW, Schmidt TFH. The ambulatory measurement of posture, thigh acceleration, and muscle tension and their relationship to heart rate. Psychophysiology. 1996;33:409-415.

[33] Bussmann JBJ, Veltink PH, Koelma F, et al. Ambulatory monitoring of mobility-related activities: the initial phase of the development of an activity monitor. Eur J Phys Med Rehabil. 1995;5:2-7.

[34] Veltink PH, Bussmann HBJ, de Vries W, et al. Detection of postures and movements using uni-axial accelerometers, IEEE Rehab Eng. 1996; 4:375-385.

[35] Jain A. Martens WLJ, Mutz G, et al. Towards a comprehensive technology for recording and analysis of multiple physiological parameters within their behavioral and environmental context. In: Fahrenberg J, Myrtek M, eds. Ambulatory Assessment: Computer-Assisted Psychological and Psychophysiological Methods in Monitoring and Field Studies. Seattle, Wash: Hogrefe & Huber Publishers; 1996:215-236.

[36] Bussmann HBJ, Tulen JHM, van Herel ECG, Stam HJ. Quantification of physical activities by means of ambulatory accelerometry: a validation study. Psychophysiology. In press.

[37] Walker DJ, Heslop PS, Plummer CJ, et al. A continuous patient activity monitor: validation and relation to disability. Physiol Meas. 1997;18:49-59.

[38] Kiani K. Snijders CJ, Gelsema ES. Recognition of daily life motor activity classes using an artificial neural network. Arch Phys Med Rehabil. 1998;79:147-154.

[39] Grant KA, Johnson PW, Galinsky TL. Evaluation of an accelerometric activity monitor as an exposure assessment tool in ergonomic studies. Appl Occup Environ Hyg. 1995;10:461-466.

[40] Kochersberger G, McConnell E, Kuchibhatla MN, Pieper C. The reliability, validity and stability of a measure of physical activity in the elderly. Arch Phys Med Rehabil. 1996;77:793-795.

[41] Robinson T, Brooke-Wavell K, Jones P, et al. Can activity monitors be used to assess compliance with walking programmers in the elderly? Int J Rehabil Res. 1995;18:263-265.

[42] Meijer GA, Westerterp KR, Koper H, ten Hoor F. Assessment of energy expenditure by recording heart rate and body acceleration. Med. Sci Sports Exerc. 1989;21:343-347.

[43] Morell EM, Keefe FJ. The actometer: an evaluation of instrument applicability for chronic pain patients. Pain. 1988;32:265-270.

[44] Bussmann JBJ, van de Laar YM, Neeleman MP, Stam HJ, Ambulatory accelerometry to quantify motor behaviour in patients after failed back surgery: a validation study. Pain. 1998;74:153-161.

[45] Davies SW, Jordan SL, Lipkin DP, Use of limb movement sensors as indicators of the level of everyday physical activity in chronic congestive heart failure. Am J Cardiol. 1992;69:1581-1568.

HBJ Bussmann, PT, is a faculty member, Institute of Rehabilitation Medicine, Faculty of Medicine and Health Sciences, Erasmus University Rotterdam, Dr Molewaterplein 40, 3015 GD Rotterdam, the Netherlands ([e-mail]). Address all correspondence to Mr Bussmann.

PJ Reuvekamp is a human movement scientist and a consultant for Services tot Disabled People, Zwolle, the Netherlands.

PH Veltink is a faculty member, Biomedical Engineering Department, Faculty of Electrical Engineering, University of Twente, Enschede, the Netherlands.

WLJ Martens is a medical physicist and a research and development consultant in the area of physiological data processing for Phyvision, Gemert, the Netherlands.

HJ Stare, PhD, MD, is Professor, Institute of Rehabilitation Medicine, Erasmus University Rotterdam, and Department of Rehabilitation, University Hospital Rotterdam, the Netherlands.

This research was part of the EUREKA Project DYNAPORT, which was financed, in part, by the Dutch Ministry of Economic Affairs.

This article was submitted July 30, 1997, and was accepted February 24, 1998.