Feature extraction and prediction of spinal cord stimulation evoked compound action potentials in humans.

Introduction Evoked compound action potentials (ECAPs) during spinal cord stimulation (SCS) may be useful in the treatment of chronic pain as a control signal for closed-loop neuromodulation. However, considerable inter-individual variability in evoked responses requires robust methods in order to realize effective, personalized pain management. These methods include artifact removal, feature extraction, classification, and prediction. Methods We recorded ECAPs from eight participants with chronic pain undergoing an externalized trial with two percutaneous leads. The two most caudal electrodes were used for stimulation and the remaining electrodes were used for recording. Artifact-cleaned waveforms were clustered using principal component analysis (PCA) and classified using a K-Nearest Neighbors (KNN) classifier as non-ECAPs, ECAPs, or outlier (i.e. artifacts) to determine how well different features, including area under the curve (AUC) and peak-to-peak amplitude (P2P), discriminate between waveform classes. Finally, we used generalized linear mixed effects models (GLMEs) to predict evoked response features and the probability of observing artifacts or ECAPs following individual stimulation pulses for different stimulation amplitudes, pulse widths, and polarities. Results AUC was better at discriminating between ECAPs and non-ECAPs than P2P (d' = 2.44 vs d' = 2.27) while most features were good at discriminating between ECAPs and artifacts (d' > 1.5). The application of an optimal AUC threshold was then used to analyze individual ECAPs at different stimulation amplitudes, pulse widths, and polarities. Interestingly, ECAPs could be evoked using ~1.25 mA less current when using participant-specific, preferred stimulation polarities. Conversely, N1 latency consistently correlated with the location of the cathode. Conclusion We developed an automated analysis pipeline for individual ECAPs during SCS. AUC was better than the widely used P2P for characterizing evoked responses. Furthermore, our modeling results provide a method for predicting how various stimulation parameters affect SCS responses in individual participants. .