Augmenting Spline Interpolated Dementia EEG Data with Deep Learning for Improved Multi-Centric Data Harmonization

Harmonizing electroencephalography (EEG) data for dementia research is propelling forward due to a recent increase in a variety of publicly available dementia EEG datasets. Understanding dementia requires data from more than a single site, but a significant problem with applying data-driven methods to the wealth of new datasets is the heterogeneity of EEG recording systems: there are different recording resolutions, and the number of electrodes and layouts varies significantly. Spatial normalization is projecting different electrode layouts into a common electrode layout and homogenizing the data by unifying them into a common analytical space. The common method of spatially normalizing multi-centric electrode layouts is through spline interpolation. In this study, we use deep learning methods to improve upon spline interpolation for better-estimated signal construction. Our results demonstrate that deep learning models create more realistic frequency domain features, which contain some of the most studied biomarkers in dementia EEG analysis.