gms | German Medical Science

Introduction: The number of people with major depressive disorder (MDD) has risen sharply. Despite the possibility of detecting systematic changes in the brain caused by MDD, diagnosis is usually based on symptoms and can be subjectively biased [1]. Therefore, research focuses on finding objective biomarkers.

Using electroencephalography (EEG)-based biomarkers in clinical applications such as diagnosis requires not only accurate but also robust biomarkers. However, robust differentiation between MDD patients and healthy controls (HC) remains challenging, with conflicting results in the literature [2]. One possible reason for this is the use of different pre-processing steps and biomarker calculations. Therefore, before improving the diagnostic accuracy with these biomarkers, investigation, standardisation, and optimisation of processing steps are needed.

Previous work has addressed this issue using multiverse analyses to investigate different processing steps on alpha-band biomarkers [3], [4]. This work aims to build on these findings to identify combined biomarkers and to provide recommendations for processing steps.

Methods: Two datasets with eyes-closed resting-state EEG data from MDD patients (n=35/34) and HC (n=35/30) were harmonised. Based on the previous study, data was normalised subject-wise, and window length was set to 20s. The following steps and variations were included in the multiverse analysis: Aggregation (10 windows per subject, median of windows), asymmetry (absolute, normalised, logarithmic), electrode combination of asymmetry (opposite, front/back), and classification (support-vector-machine, random forest, logistic regression, multi-layer perceptron).

Two approaches were tested to find the optimal set of biomarkers. For the top-down approach, 87 biomarkers (12 spectral markers per 7 frequency bands + 2 coherences across bands + the central frequency) are analysed in 48 models using the multiverse analysis. Robust biomarkers are identified by analysing the accuracies and the feature importance of the resulting models. For the bottom-up approach, we analyse all biomarkers in their multiverses individually (87*48 = 4176 models) and subsequently combine the most robust and powerful individual markers to the optimal combined biomarker.

Results: The accuracies of the 48 top-down models range between 86% and 93%. The best-performing set was a combination of median of windows, the logarithmic asymmetry with opposite electrodes, and a logistic regression as classification algorithm. In the bottom-up approach, the best individual spectral biomarker was the absolute centroid frequency of the gamma band (88.9±1.0%), closely followed by the relative centroid frequency in the same band (88.0±0.8%). The results of the feature importance analyses and the combined biomarkers will be presented at the conference.

Discussion: Building on the previous work, the robustness and accuracy for clinical use can be further improved by adding biomarkers from other frequency bands. The preliminary results of this study have already indicated promising variations in the processing steps for the robust use of EEG biomarkers in the clinical diagnosis of MDD. These results will be further refined by combining individual biomarkers.

Conclusion: The results of the study show that differences in processing steps have a major impact on the accuracy of the classification models. Accurate information on the processing steps and their standardisation, will further advance the clinical use of EEG biomarkers.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.