30756
Spline Fitting for More Robust ERP Derived Dependent Variables
Electroencephalography (EEG) is a valuable tool for studying Autism Spectrum Disorder (ASD) due to its high temporal resolution. By repeated exposure of participants to stimuli, researchers can construct Event Related Potentials (ERPs). For statistical analysis, measures of ERP peaks often utilize automated window-based peak picking (AWPP) to find peak activity during a specified time window. AWPP offers no reliable way to assess accuracy of derived dependent variables (DVs), such as peak amplitude and latency to peak. Manual peak picking (MPP), in which humans select peaks based on visual inspection, offers this benefit but is significantly more time consuming and prone to human error. Splines are smooth lines made of Bezier curves. Fitting splines to ERPs presents an alternative method to derive DVs, which may offer a middle ground of performance assessment and time efficiency. Given the unusual waveform morphologies commonly observed in ASD, this would represent a significant advance, particularly in large samples.
Objectives:
We sought to: (1) develop an algorithm to fit splines to ERP data; (2) assess the success in both ASD and typical development (TD) using spline parameters instead of AWPP; (3) identify avenues for further development as a tool for analyzing EEG data in ASD.
Methods:
Data were collected across 106 EEG sessions with adult participants clinically diagnosed with ASD or TD controls. Participants were shown dynamic faces displaying emotional expressions. Data were processed with simple filtering and artifact detection and averaged across trials. AWPP was used to find a positive peak within a window of 40-190ms relative to the event (P100), and a negative peak in the window of 120-250ms (N170), both in averaged channel groups representing the left and right occipto-temporal scalp. A grand average was used to manually create a starting spline, which was then automatically fit to each ERP for the same channel groups by our algorithm. The parameters that defined the fit splines were used in analysis.
Results:
Based on simple ANOVAs using both AWPP values and corresponding spline control points values (SCPV), our algorithm was successful in extracting meaningful data from ERPs. A statistically significant difference between groups was detecting using both AWPP-derived N170 amplitude (p=.015) and the SCPV (p=.007). In addition to comparable performance in group discrimination, the SCPV offer a goodness of fit with R2 values, and thus an estimation of divergence from recorded waveforms. Further, some SCPVs with no AWPP equivalent show promise of discriminatory power, indicating that our method may capture additional, novel aspects of waveform shape.
Conclusions:
Our promising results warrant further development of these methods. Given the same data, novel methods were able to extract equivalent meaningful information from the ERPs when compared to AWPP. Further, goodness of fit estimation and novel values with possible discriminatory power offer advantages over the traditional method. While our methods were not compared to MPP, the vastly lower cost in human effort highlights the value of this approach in quantifying individual differences in an automated and unbiased fashion. Ongoing analysis is examining our approach with other ERPs, such as Visual Evoked Potentials.