4.4 Single-sensor System
4.4.3 Results
Kz,hem= αhem· Pref
Psum+ βhem (4.7)
Kz,nor= αnor· Pref
Psum+ βnor (4.8)
where: the ratio between reference and subject’s spectrum powers is the same defined in (4.2); and the linear models’ coefficients ({αhem, βhem} and {αnor, βnor} for paretic and non-paretic sides, respectively) are empirically defined in a preliminary tuning phase (according to the experimental evidences) and remain fixed across all subjects.
The choice behind these values is motivated by the intuition that the relationship be-tween COM vertical and horizontal displacements, considering a step of the affected side, is weakened by the hemiparetic patients’ tendency to rise the paretic-side hip during the swing phase, as a consequence of a compensatory movement needed to complete the forward progression[112]. This behavior introduces “noisy” COM ver-tical movements. Using a lower value of Kzfor the paretic side allows to reduce this phenomenon and to estimate more precisely, the step lengths of both sides.
For what concerns the extended version of the Weinberg algorithm, the only modification concerns the subject-dependent calibration constant, denoted as Kw,mod, which is automatically calculated with the following two-variable linear model, con-sidering both Psumand the leg length `:
Kw,mod= αw,P· Pref
Psum+ αw,`· ` + βw (4.9) where the model coefficients {αw,P, αw,`, βw}, as for (4.7) and (4.8) cases, are em-pirically defined (just once) to achieve the best fit with the actual step lengths across different individuals. Once all the step lengths have been computed with the proposed modified Zijlstra and Weinberg algorithms, stride lengths (StrideLR/L) and velocities (StrideVR/L) can be easily calculated.
4.4. Single-sensor System 135
Table 4.2: Temporal parameters estimation results.
Post Stroke patients Healthy Subjects Feature [dimension] Mean±Std RMSE R Mean±Std RMSE R
Psum[adim.] 2352 ± 717 - - 3858 ± 278 -
-M[samples] 22.8 ± 6.47 - - 15.5 ± 3.44 -
-GCTR[s] 1.45 ± 0.11 0.03 0.97 1.27 ± 0.19 0.02 0.99 GCTL[s] 1.47 ± 0.13 0.04 0.96 1.30 ± 0.21 0.02 1.00 STR[% of GCT] 69.3 ± 7.2 2.60 0.97 63.2 ± 1.6 2.02 0.99 STL[% of GCT] 67.8 ± 6.7 2.89 0.94 64.1 ± 2.4 1.91 0.99 SWR[% of GCT] 30.7 ± 7.2 2.60 0.94 36.8 ± 1.6 2.02 0.90 SWL[% of GCT] 32.2 ± 6.7 2.89 0.92 35.9 ± 2.4 1.91 0.93 IDS[% of GCT] 18.9 ± 3.5 4.02 0.57 13.1 ± 1.5 1.25 0.92 T DS[% of GCT] 18.5 ± 3.5 3.74 0.64 14.9 ± 3.1 1.97 0.94 DS[% of GCT] 37.4 ± 4.4 4.01 0.80 28.2 ± 4.2 2.48 0.96 Cstep[steps/min] 82.1 ± 5.6 1.02 0.99 98.3 ± 11.8 0.63 1.00
with those estimated by the IMU-based approaches. The Root Mean Square Error (RMSE) and the Pearson’s correlation coefficient, denoted as R, between the features measured with the two methods are considered as performance metrics.
In Table 4.2, the main results, in terms of temporal parameters’ estimation, ob-tained with the proposed algorithm are shown. In the first two rows, the average val-ues of the main parameters used in the initial step detection procedure, namely Psum
and the SMA window size M, are shown, for both post-stroke and healthy subjects. Is it possible to observe that the average Psumin healthy subjects is considerably higher than in hemiparetic patients, highlighting the effectiveness of this concise parameter in representing the overall level of gait impairments. The temporal parameter val-ues (mean±standard deviation), estimated using the power-dependent SMA window size, are shown in the lower part of Table 4.2, together with the performance met-rics RMSE and R. As expected, the estimated hemiparetic gait characteristics reflect irregular and asymmetric walking patterns, according to the clinical observations.
Post-stroke subjects, indeed, show higher GCT , ST , and DS values and significantly lower step cadence than controls. Moreover, single patient analyses, which are not
Table 4.3: Spatial parameters estimation results.
Post Stroke patients Healthy Subjects Feature [dimension] Mean±Std RMSE R Mean±Std RMSE R
Kz,hem 0.66 ± 0.08 - - - -
-Kz,nor 0.87 ± 0.12 - - - -
-Kw,mod 0.25 ± 0.04 - - 0.47 ± 0.04 -
-StepLz,R[m] 0.32 ± 0.08 0.03 0.92 - -
-StepLz,L[m] 0.35 ± 0.08 0.03 0.93 - -
-StepLw,R[m] 0.33 ± 0.09 0.04 0.89 0.69 ± 0.07 0.05 0.81 StepLw,L[m] 0.34 ± 0.08 0.04 0.91 0.70 ± 0.06 0.06 0.82
StrideVz,R[m/s] 0.45 ± 0.12 0.03 0.97 - -
-StrideVz,L[m/s] 0.43 ± 0.10 0.03 0.95 - -
-StrideVw,R[m/s] 0.46 ± 0.13 0.03 0.97 1.12 ± 0.22 0.08 0.98 StrideVw,L[m/s] 0.43 ± 0.11 0.03 0.95 1.06 ± 0.22 0.07 0.98
shown here, highlight a consistent increasing in the non-paretic limb ST feature, as a consequence of the gait asymmetry [112]. Comparing the performance of the pro-posed system with that of the BTS mo-cap system, the RMSE ranges between 20 ms and 40 ms for the GCT , from 1.25% to 4.02% of the GCT for stride phase param-eters, and it is approximately equal to 1 step/min for the step cadence feature. The correlation coefficient between reference and estimated features’ values is very high (R > 0.90) for all the parameters, except those of post-stroke patients’ double support phases (namely, IDS, T DS, and DS), which range from 0.5 to 0.8.
For what concerns the spatial parameters, we first present the coefficients’ values of the linear models used for the calibration constants’ automatic computation, as de-scribed in Subsection 4.4.2. For the modified Zijlstra algorithm, the values of Kz,hem in (4.7) are obtained with αhem= −0.04 and βhem= 0.75, whereas Kz,nor in (4.8) is computed using αnor= −0.085 and βnor= 1.03. The two-variable linear regression model (4.9) used for computing Kw,modin both post-stroke and healthy subjects uses the following coefficients: αw,P= 0.006, αw,`= 2.31, and βw= −1.85. In the first two rows of Table 4.3, the average values of the proposed step length estimators’ calibra-tion constants are shown. In the remaining rows of Table 4.3, step length and stride
4.4. Single-sensor System 137
velocity values, computed using the modified versions of both Zijlstra and Weinberg algorithms, are shown. As reported by several clinic studies [134], step/stride length and velocity are crucial features to assess the walking ability level of post-stroke in-dividuals. It can be observed, in fact, that the average value of all the hemiparetic features is approximately doubled in healthy subjects. For post-stroke patients, the RMSEs of StepL and StrideV are around 3 cm and 3 cm/s, respectively, considering both the modified Zijlstra and Weinberg algorithms. For the healthy group of subjects, the asymmetric version of Zijlstra step length estimator is not applicable, since their gait patterns are symmetric, whereas the StepL and StrideV RMSEs obtained using the modified Weinberg method are approximately equal to 5 cm and 7 cm/s, respec-tively. As for the temporal parameters, the correlation coefficient values is very high (R ≥ 0.89) for all the spatial features of both groups, except for the StepLw,R/L val-ues of healthy subjects, for which the correlation coefficient is slightly lower (around 0.82).
The high correlation between reference and estimated spatio-temporal parame-ters is clearly shown in Figure 4.8. Most of the points, whose(x, y) coordinates are the values measured by the optoelectronic and the IMU-based systems, respectively, are located near the dotted diagonal line, which represents the ideal case with per-fect correlation. As previously remarked, DS for post-stroke patients and StepLwfor healthy subjects are the parameters which show the lowest correlations with the cor-responding reference values. This is due to the fact that, in hemiparetic individuals the detection of the actual T O instants can be missed in some cases, anticipating or delaying it of a few tens of milliseconds. At the opposite, for the healthy subjects the regression model for the computation of Kw is slightly polarized toward the correct estimation of step lengths in the hemiparetic case, reducing the accuracy for normal walking subjects.
Finally, an overall view on the accuracy achieved by the proposed methods is given in Figure 4.9, visualizing with box plots the characteristics (median, quartiles and extreme values) of the most relevant estimated spatio-temporal parameters and comparing them with the features extracted from the reference mo-cap system – both post-stroke and healthy subjects’ values are included in the computation of each box
Reference System [s]
0.8 1 1.2 1.4 1.6 1.8 2
IMU-basedSystem[s]
0.8 1 1.2 1.4 1.6 1.8
2 GCTR/L(post stroke) GCTR/L(healthy)
-(a) GCT
Reference System [% of GCT]
0 20 40 60 80 100
IMU-basedSystem[%ofGCT]
0 10 20 30 40 50 60 70 80 90
100 STR/L(post stroke) DS (post stroke) STR/L(healthy) DS (healthy)
-(b) ST and DS
Reference System [m]
0 0.2 0.4 0.6 0.8 1
IMU-basedSystem[m]
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
1 StepLz,R/L(post stroke) StepLw,R/L(post stroke) StepLw,R/L(healthy)
-(c) StepL
Reference System [m/s]
0 0.5 1 1.5 2
IMU-basedSystem[m/s]
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8
2 StrideVz,R/L(post stroke) StrideVw,R/L(post stroke) StrideVw,R/L(healthy)
-(d) StrideV
Figure 4.8: Comparison between features values calculated through the optoelectronic reference system and those estimated with the proposed IMU-based methods for post-stroke (red markers) and healthy (blue markers) subjects: (a) GCT , (b) ST and DS, (c) StepL, and (d) StrideV .
4.4. Single-sensor System 139
GCTref,R/L GCTR/L STref,R/L STR/L DSref DS
NormalizedValue
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
(a) Temporal parameters
StepLref,R/L StepLz,R/L StepLw,R/L StrideVref,R/L StrideVz,R/L StrideVw,R/L
NormalizedValue
-1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
(b) Spatial parameters
Figure 4.9: Boxplots of some representative (a) temporal and (b) spatial parameters.
plot. Once again, it can be observed that the similarity between reference and es-timated values is very high, with a slight decreasing in accuracy only for DS ans StepLwfeatures.
The obtained results in both temporal and spatial parameters extraction are con-sistent with other works in the literature focused on hemiparetic GA of post-stroke patients [135, 123]. Moreover, the achieved accuracy compares favorably with other general-purpose IMU-based GA algorithms, for both pathological and normal walk-ing assessment [59, 55].