Predictive modeling to assess the effectiveness of epidural stimulation parameters that promote standing in individuals with severe spinal cord injury

Introduction: Data mining techniques have growing applications in large datasets in healthcare to enable researchers and healthcare professionals to systematically use machine learning tools to identify patterns in the data and use it to improve protocols and predict future outcomes. We have been collecting EMG activity data during standing with spinal cord epidural stimulation for a decade on individuals with motor complete spinal cord injury (SCI). It is also a well-known fact that finding proper stimulation parameters (i.e. electrode combination, polarity, intensity and frequency) is a key factor in promoting independent standing with epidural stimulation. In this study, we have developed a predictive modeling framework to address the challenging task of stimulation parameters selection that leads to independent standing. Materials and Methods: Eleven individuals with chronic, clinically motor complete or sensory and motor complete SCI individuals are included in this study. The EMG signals were recorded from 16 proximal and distal leg muscles during the performance of standing task with epidural stimulation. We proposed a predictive modeling framework (Fig. 1) that uses spectral features of EMG and multiple classifiers to find neurophysiological patterns in the EMG recordings from 16 leg muscles that lead to independent standing. This framework then uses the trained models to predict the effectiveness of each set of stimulation parameters for promoting independent standing performance in 48 assisted standing events performed for 6 participants while different scES parameters were tested. Results and Discussion: We have shown that the trained KNN classifiers can perform with average accuracy of 96% to discriminate assisted standing from independent standing conditions. The proposed prediction algorithm can also score the performance of each investigated muscle based on the trained patterns when different stimulation parameters are selected (Fig. 2). The results of the prediction algorithm suggests that the proposed framework can provide reliable feedback regarding which stimulation parameters can improve the performance of each muscle that would subsequently lead to independent standing. Conclusion: The proposed framework in this study can fast-track the search for proper sets of stimulation parameters, and therefore improve standing motor recovery in the SCI population.