gms | German Medical Science

Introduction: The neural signature of Major Depressive Disorder (MDD) obtained with Electroencephalography (EEG) is still unclear. Results vary widely regarding the selectivity of features extracted from EEG signals. For instance, alpha bandpower has been identified as a promising biomarker for EEG-based MDD diagnosis with conflicting results. In some studies, it has higher diagnostic power than bandpower from other EEG bands, in other studies less [1]. Problems arise from small datasets based on which study results lack generalization and replicability. Additionally, differences in data recording, pre-processing, feature extraction, and classification algorithms hamper study comparison. In order to apply EEG-based biomarkers in clinical use cases like diagnosis, robust biomarkers are needed. This study addresses the problems by joining two datasets to systematically investigate with dataset harmonization combined with multiverse analysis [2] the effects of methodological differences.

Methods: Two public datasets with eyes-closed resting-state EEG data from MDD patients (n=24/34) and HC (n=29/30) [3], [4] were harmonized by selecting 13 corresponding EEG channels, re-referencing to average signal, resampling to 250Hz, and band-pass filtering (1Hz/40Hz). Artifacts were automatically labelled and removed [5]. With the subsequent pre-processing steps, we conducted the multiverse analysis: Normalization (none, subject-wise, channel-wise), window splitting (5, 10, 15, 20 sec), outlier removal, selection of 10 random windows per subject, alpha bandpower calculation per window, and augmentation (10 windows per subject, median of windows) were applied to three datasets (dataset 1, dataset 2, combined dataset). To investigate the robustness, we trained 72 “MDD diagnosis classifiers” (linear Support Vector Machine) resulting from three datasets * 24 pre-processing conditions. To investigate possible differences between datasets, we trained 24 “dataset separation classifiers” on the combined dataset.

Results: For the diagnosis classifiers, we found no effects of classification accuracy on window length (F(3, 720)=1.885; p=.131), and therefore constrained further analyses to 20 sec windows. Normalization, however, had a sign. effect on accuracy (F(2,180)=13.522; p<.001) with channel-wise normalization yielding the lowest accuracy. The 18 remaining parameter combinations yielded highly inconsistent results for classification, exceeding chancel level in only 8 cases (p<.05). Merely one survived the correction for multiple comparisons. In contrast, the two datasets were found separable in all cases except when channel-wise normalization was combined with the median for augmentation (t(9)=0.638; p=.724).

Discussion: Our study mirrors the inconsistencies found in the literature for alpha bandpower as MDD biomarker with controllable variations. Datasets used, i.e. subject selection and data recording, and variation of commonly used pre-processing steps strongly influence this biomarker's diagnostic capability. Furthermore, pre-processing methods that equalize differences and facilitate comparability between datasets negatively affect the diagnosis. However, alpha bandpower is only one possible MDD biomarker and further EEG features as well as their combination provide further opportunities for obtaining robust biomarkers.

Conclusion: We introduce a multiverse analysis that compares different commonly used variations of pre-processing and datasets. Results were inconsistent across datasets and pre-processing methods, highlighting the sensitivity of EEG biomarkers to such variations. We therefore advocate for further methodological studies into robust processing of EEG data to obtain biomarkers ready for clinical routine.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.