Table 1.
Descriptions of the extracted gait parameters.
Table 2.
Hyperparameters obtained for the classifiers.
Fig 1.
ROC curves of the adopted ML models.
Table 3.
Performance of classifiers on test data.
Fig 2.
Feature importance for the RF and LightGBM models.
Fig 3.
Global interpretability for (a) RF, (b) SVM, (c) LR, and (d) LightGBM. The blue dots and red dots on the right side of the vertical line indicate low and high values, respectively, of corresponding gait parameters that have an influence on ATK. Similarly, the dots on the left side of the vertical line are for ICRC. Features at the top are of the most importance. For example, for Prosthetic Flex SI, the ATK has lower values (i.e. less prosthetic knee flexion) compared to ICRC. The changes in the gait parameters are due to the ATK swing-phase control mechanism. The ATK mechanism better controls (compared to ICRC) the kinematics at the knee joint during gait to increase symmetry between the prosthetic and non-prosthetic sides, specifically related to swing times, step lengths as well as heel-rise (swing-phase knee flexion excursions) across different walking speeds. For a more detailed explanation please refer to [3].
Table 4.
RFE results for the ML models.
Fig 4.
SHAP dependence plots for four most influential decreasing gait parameters via RF model.
For the low values of prosthetic max knee flex (up to 73 degrees), the likelihood of wearing an ATK was high. After passing 73 degrees, the probability of wearing an ATK started to decline meaning that the ICRC knee was likely worn. A similar pattern was observed for knee flex SI, swing time SI, and prosthetic swing time with the critical points of roughly 7, 33, and 0.52 sec, respectively.
Fig 5.
SHAP dependence plots for gait parameters ranked five to eight for the RF model.
In general, all four gait parameters started from low values for ICRC knees and increased for the ATK knees. However, the trend and rate of the curves are different for different features. For example, stride velocity ankle shows a flat trend for the values between 0.9 m/sec and 1.1 m/sec.
Fig 6.
SHAP dependence plots for four least influential gait parameters via RF model.
The general trend for the dependence plots of these features was flat indicating that these gait parameters had minimal impact on the model predictions. This means that neither ATK nor ICRC prosthetic knees affected these gait parameters. The dependence plot for prosthetic step length parameter demonstrates a nonlinear pattern between its varying values and the likelihood of wearing the prosthetic knees implying both ATK or ICRC knees could result in either low or high values of prosthetic step length (but the impact is low in any case).
Fig 7.
Local interpretability analysis for three samples using RF model.
The red and blue arrows correspond to gait parameter values that push the model toward the ATK and ICRC knee class, respectively. The amount of variation of each gait parameter is determined by the length of the arrow (or the value associated with it). P1, P2, and P3 refer to three randomly selected study participants.
Fig 8.
Local interpretability analysis for two samples wearing ATK prosthesis using the four models.
P1 and P2 are randomly selected study participants.
Fig 9.
SHAP analysis for a false positive example of classification for a single participant wearing an ICRC prosthesis obtained from the SVM model.
Since the label -1 was assigned to the ICRC, one expected to see a vertical line labelled with -1. However, the class label was marked +1, meaning that the model had identified this sample with an ATK prosthesis.