This section evaluates the agreement and reliability of the developed system on its two configurations (C1 and C2) compared with a reference (manual) method. In addition, the performance of both system configurations is also compared.
Plotly loading issue
This page contains Plotly interactive figures. Sometimes, the figures might not load properly and show a blank image. Reloading the page might solve the loading issue.
In some ocassions, although the systems are able to obtain valid results, the way these results are obtained is incorrect. For instance, the system might not be able to correctly detect when the subject has sit down (due to trousers/pocket characteristics), but ending up detecting the end of the test after the subject extracts the smartphone from the pocket and generating apparently valid results. Therefore, these results are not valid and must be marked as invalid. Following, these cases are reported:
s10: all the results from the smartphone (C2) are invalid. The pocket of the subject was small, with half of the smartphone being outside it. When the subject sat down, the smartphone was in a position that the system was not able to recognize as SEATED.
s11: the \(4^{th}\) result from the smartphone (C2) is invalid. The subject wore pants with loose pockets, and in the \(4^{th}\) execution was not able to detect the subject as being SEATED.
s25: the \(3^{rd}\) and \(5^{th}\) results from the smartphone (C2) are invalid. The pocket of the subject was small, with half of the smartphone being outside it. When the subject sat down, the smartphone was in a position that the system was not able to recognize as SEATED.
Next, these cases are manually modified in order to invalidate them (i.e., setting their duration to -1).
Analyzing C1 (smartwatch) results
Found partial success: 19
Found partial success: 23
Found partial success: 24
Found partial success: 26
Found failure: 41
Found failure: 43
Found partial success: 80
Found failure: 81
Found partial success: 84
Found failure: 104
Found partial success: 110
Found partial success: 155
Found partial success: 163
Found partial success: 164
Found partial success: 192
Found failure: 201
Found partial success: 207
Found partial success: 209
Found partial success: 243
Found partial success: 244
Found partial success: 248
Found failure: 275
Analyzing C2 (smartphone) results
Found failure: 10
Found failure: 11
Found failure: 14
Found failure: 16
Found failure: 18
Found failure: 43
Found failure: 55
Found failure: 89
Found failure: 90
Found failure: 91
Found failure: 92
Found failure: 93
Found failure: 94
Found failure: 95
Found failure: 96
Found failure: 97
Found failure: 98
Found failure: 102
Found failure: 112
Found partial success: 120
Found failure: 130
Found partial success: 132
Found partial success: 135
Found failure: 138
Found failure: 139
Found failure: 140
Found failure: 141
Found failure: 146
Found partial success: 154
Found failure: 191
Found partial success: 197
Found failure: 200
Found failure: 201
Found failure: 204
Found partial success: 210
Found failure: 223
Found failure: 229
Found failure: 231
Found failure: 232
Found failure: 233
Found failure: 234
Found failure: 235
Found failure: 236
Found failure: 237
Found failure: 238
Found failure: 239
Found failure: 240
Found failure: 241
Found failure: 264
Found failure: 266
Found failure: 267
Found failure: 268
Found failure: 269
Found failure: 270
Found partial success: 282
Table 10.2 shows the success, partial success and failure results classification. C1 obtained a higher success and partial success rate than C2 (\(92.44\%\) and \(5.5\%\) vs. \(81.1\%\) and \(2.41\%\)). Regarding failures, C1 obtained a better rate than C2 (\(2.06\%\) vs \(16.49\%\)).
A deeper analysis of the failures showed that male participants were responsible for \(5\) out of the \(6\)failures in C1, where the system failed to recognize the standing_up subphase. On the other hand, female participants were responsible for \(39\) out of the \(48\)failures in C2, mainly due to wearing jeans with small pockets that placed the smartphone in an un-favourable position, which significantly affected the recognition of the sitting_down subphase.
Table 10.2: Classified results obtained by the system in both configurations.
status
failure
partial_success
success
system
C1
6
16
269
C2
48
7
236
Figure 10.1 shows the distribution of the measurement error (i.e., system_measure\(-\)manual_measure) for the total duration (i.e., duration) of the test and each subphase (i.e., standing_up, first_walk, first_turn, second_walk, second_turn and sitting_down) for the executions classified as success. In addition, the mean of the distribution of each measurement is compared to \(0\) with a T-test or a W-test. Results indicate that the mean of the measurement errors of standing_up and sitting_down are not significantly different from \(0\) in both configurations, neither are the errors of first_turn and duration in C1 and C2, respectively. These results indicate that, on average, the system provides very accurate measurements for the mentioned metrics and configurations compared with the reference measurements.
Code
error_distribution(errors_df)
Figure 10.1: Measurements error’s distribution for the total duration and each subphase for success executions in C1 and C2. At the bottom are the p-values of the one-sample T-test (green) or W-test (white) comparing the means of each distribution with \(0\), where a p-value > \(0.05\) indicates no significant differences between the distribution mean and \(0\).
Table 10.3 shows the inter-subject RMSE and the results of the statistical tests executed comparing each measure in both systems. C1 reported significantly better results in the duration, first_turn and sitting_down measures. Other not significant differences were observed in standing_up and second_walk in favour of C1, and in first_walk and second_turn in favour of C2. These results denote that the system performs better in C1 than in C2 when taking into account different subjects (i.e., C1 is more adaptable to different subjects).
Code
compare_rmse_distributions(errors_df)
Table 10.3: Comparison of the inter-subject RMSE of each measurement for both system configurations.
Measure
M(C1)
M(C2)
Test
0
duration
287.000000
389.000000
U=280.5, p-val=0.031, power=0.542
1
standing_up
286.766667
336.821429
t(55.90)=-1.24, p-val=0.222, power=0.227200896...
2
first_walk
296.000000
268.500000
U=460.5, p-val=0.534, power=0.338
3
first_turn
278.500000
356.000000
U=285.5, p-val=0.037, power=0.434
4
second_walk
291.000000
338.000000
U=326.0, p-val=0.146, power=0.283
5
second_turn
233.000000
217.000000
U=410.5, p-val=0.889, power=0.197
6
sitting_down
269.500000
322.500000
U=280.0, p-val=0.03, power=0.603
Bland-Altman analysis
To determine the agreement between each system configuration and the reference measures, Figure 10.2 and Figure 10.3 show the Bland-Altman agreement analysis for the duration and the rest of the measures, respectively. For the duration measurement, C1 and C2 show an excellent mean of \(-0.044\) and \(-0.051\) with \(95\%\) agreement limits of \([-0.7,0.61]\) and \([-0.91,0.8]\). These results indicate a low bias with regards to the reference method and that \(5\%\) of the measurements exceed an error of \(0.7\) seconds in C1 and \(0.9\) in C2.
Regarding the measures of the TUG subphases, C1 shows slightly better means (all under \(0.1\) seconds) than C2 in all subphases except for the standing_up. The limits of agreement in C1 are around \(0.6-0.9\) and in C2 exceed the second. C1 results indicate a low bias with respect to reference measures, while C2 shows biases of around \(0.2\) seconds for walking and turning activities, which are underestimated and overestimated, respectively.
Figure 10.2: Blant-Altman analysis of the total duration of the test computed by the system in both configurations. The red line represents the mean difference and both dotted lines are the limits of agreement (with \(95\)% CI).
Figure 10.3: Blant-Altman analysis of the TUG subphases of the test computed by the system in both configurations. The red line represents the mean difference and both dotted lines are the limits of agreement (with \(95\)% CI).
Intraclass Correlation Coefficent (ICC) analysis
To measure the reliability of both configurations, Table 10.4 contains the ICC estimates and their \(95\%\) confident intervals based on a single rater, absolute-agreement, 2-way mixed-effects model comparing the measures obtained by both system configurations with the reference ones. ICC results show excellent reliability for the duration in both system configurations. Regarding the subphases, C1 achieves excellent reliability in the second_walk; good reliability in first_walk; moderate reliability in second_turn and sitting_down, and poor reliability for the remaining ones, whereby the reliability of the first_turn is non-significant (i.e., CI contains 0). C2 achieves moderate reliability in first_walk and second_walk while obtaining poor reliability in the rest, with non-significant results for standing_up, first_turn and second_turn.
Table 10.4: Intraclass correlation from single rater, absolute-agreement, 2-way mixed-effects model between system’s configurations measurements and the reference (manual) measurements.
ICC
CI
F Test
p-value
phase
system
duration
C1
0.979
[0.96, 0.99]
94.521
0.000
C2
0.972
[0.94, 0.99]
72.498
0.000
standing_up
C1
0.456
[0.12, 0.7]
2.677
0.005
C2
0.176
[-0.22, 0.52]
1.414
0.187
first_walk
C1
0.822
[0.63, 0.91]
11.755
0.000
C2
0.687
[0.32, 0.86]
7.076
0.000
first_turn
C1
0.264
[-0.11, 0.57]
1.694
0.081
C2
0.026
[-0.19, 0.3]
1.085
0.417
second_walk
C1
0.916
[0.83, 0.96]
24.004
0.000
C2
0.627
[0.1, 0.84]
6.922
0.000
second_turn
C1
0.540
[0.2, 0.76]
3.987
0.000
C2
0.107
[-0.19, 0.42]
1.298
0.251
sitting_down
C1
0.697
[0.45, 0.84]
5.471
0.000
C2
0.454
[0.1, 0.7]
2.634
0.007
Summary
These results (errors, Bland-Altman and ICC analyses) show that while the system is reliable in terms of measuring the total duration of the test, measurements of individual subphases must be considered carefully, especially when using the system with configuration C2. This is probably caused by the underestimation of some subphases while overestimating others, leading to a reliable total duration but less reliable subphases’ durations.
When comparing C1 and C2 configurations, the presented results provide evidence that the smartwatch-based system performs better than the smartphone-based system, both for the overall test duration and for the individual subphases detection.
Overall, the developed system, using off-the-shelf hardware (i.e., smartphone and smartwatch), presented state-of-the-art results in the computation of the TUG test duration, even when compared with systems using specialized sensor units and/or positioning sensing devices in unnatural positions (e.g., strapped to the arms or legs). This paves the way to real-life applicability (e.g., self-administration) of the TUG test and other similar tests. On the other hand, the computed duration of the TUG subphases must be considered carefully, especially in the C2 configuration, which is penalized given the diversity of front pockets and phone orientations.
Code reference
Tip
The documentation of the Python functions employed in this section can be found in Chapter 4 reference: