Automated Labeling of Movement- Related Cortical Potentials Using Segmented Regression

The movement-related cortical potential (MRCP) is a brain signal related to planning and execution of motor tasks. From an MRCP, three notable features can be identified: the early Bereitschaftspotential (BP1), the late Bereitschaftspotential (BP2), and the negative peak (PN). These features have been used in past studies to quantify neurophysiological changes in response to motor training. Currently, either manual labeling or a priori specification of time points is used to extract these features. The limitation of these methods is the inability to fully model the features. This paper proposes the segmented regression along with a local peak method for automated labeling of the features. The proposed method derives the onsets, amplitudes at onsets, and slopes of BP1 and BP2 along with time and amplitude of the PN in a typical average MRCP. To choose the most suitable regression technique a bounded segmented regression method, a change point method and multivariate adaptive regression splines were evaluated using the root-mean-square error on a dataset of 6000 simulated MRCPs. The best-performing regression technique combined with the local peak method was then applied to a smaller set of 123 simulated MRCPs. Error in onsets of BP1 and BP2 and time of PN were compared with the errors in manual labeling by an expert. The performance of the proposed method was also evaluated on an experimental dataset of MRCPs derived from electroencephalography (EEG) recorded across two sessions from 22 healthy participants during a lower limb task. The Bland-Altman plots were used to evaluate the absolute reliability of the proposed method. On experimental data, the proposed method was also compared with manual labeling by an expert. Bounded segmented regression produced the smallest error on the simulation data. For the experimental data, our proposed method did not exhibit statistically significant bias in any of the modeled features. Furthermore, its performance was comparable to manual labeling by experts. We conclude that the proposed method can be used to automatically obtain robust estimates for the MRCP features with known measurement error.