Estimating knee moments from plantar pressure data in oa subjects

Abstract

Purpose: The knee adduction moment (KAM) has been linked to the severity and progression of knee OA. 3D motion capture (MOCAP) with cameras and force plates provides joint kinematics and kinetics, and is the gold standard in determining KAM and other knee moments. While useful for research, this technology is restrictive in the clinical setting due to space and time limitations. Previously, artificial neural networks (ANN) have been used to estimate knee kinematics from portable inertial measurement units. The possibility to predict joint kinetics from plantar pressure data, which are easily obtainable from portable pressure platforms or shoe insoles, would allow data collection in the doctor’s office or in the habitual environment of the patient. Here, we sought to process standard output data from a clinically applicable plantar pressure platform to generate accurate estimates of KAM and the knee flexion moment (KFM) by the use of a novel ANN network application. We hypothesized that KAM and KFM can be predicted using plantar pressure data in combination with ANN. Methods: To estimate the KAM and KFM with the pressure platform, this IRB approved study evaluated data from patients with mild to moderate knee OA. Twenty-eight subjects (age: 59.32 ± 7.58 years; height: 1.66 ± 0.09 m; weight: 81.75 ± 13.4 kg: gender: 21F, 7M) completed three visits with time intervals of 12 week between visits. All participants underwent 3D MOCAP while performing five normal barefoot walking trials at self-selected speed. Simultaneous force and pedobarographic data were acquired at 100 Hz using a pressure-detecting platform (Novel, Munich Germany) stacked on a force plate (Bertec, Columbus, OH) and leveled with the walkway. KAM and KFM calculated from MOCAP data was then compared with the KAM and KFM estimates using ANN modelling of plantar pressure and force data generated from the pressure platform only. Two types of ANNs were evaluated: a classical feed forward artificial neural network FFANN and a long short-term memory neural network (LSTM). The LSTM incorporates units which allow the model to recall information from previous iterations and learn long-term dependencies from the parameters. For the FFANN several architectures with one or two hidden layers of varying amounts of neurons were assessed; the FFANN was trained with individual and combined standard parameter outputs from the pressure platform software. Of primary interest was the medial/lateral center of pressure (COP) location, maximum value pictures (MPP) and force. Afterwards, the input variable that lead to better performance was used as input parameter for the LSTM; this network was composed of two stacked layers with 256 and 124 units. Finally, second and third visit moment predictions were executed with the best performing network. Pearson correlation coefficients and normalized root mean square errors served as evaluation metrics. Results: In general, tested input variables showed a moderate degree of correlation with KAM and KFM MOCAP data. Table 1 shows the evaluation metrics for several individual and combined pressure data inputs for FFANN models. Individual input parameters were outperformed by combinations of them. The best results were achieved with a combination of force vector and pressure pictures. Moreover, the LSTM network outperformed the FFANN; in Figure 1, it is noticeable how LSTM reduced the gap between the MOCAP and the predicted results, especially for the KFM, in which abrupt changes were more evident than with the FFANN. Accordingly, the performance metrics using the LSTM network increased (Table 2). When using the LSTM network to predict moments for the 12- and 24-weeks visits, all the metrics demonstrated stable behavior, with similar values as during baseline. Conclusions: Based on the NRMSE and R- values obtained in this study, the model can predict KAM and KFM with moderate accuracy. This suggests that clinically useful knee moments can be predicted from plantar pressure data when combined with appropriate ANN models. In order to overcome the uncertain time alignment between input variables and knee moments, we applied LSTM modeling, which clearly is advantageous over the classical FFANN approach. The study itself relied on a relatively small (although gait-study typical) data set of n=28. This may have limited the model’s learning capacity; especially, in the presence of clinically relevant outliers. However, our model should improve over time as more data from the pressure platform and MOCAP are imported. In addition, anthropometric variables (e.g. valgus/varus alignment), which do not require advanced gait lab settings could be incorporated into the model to improve accuracy. In the future we hope to apply this technology to capture biomechanical markers in the field.View Large Image Figure ViewerDownload Hi-res image Download (PPT)View Large Image Figure ViewerDownload Hi-res image Download (PPT)