Detection of event-related potentials as a biomarker in major depressive disorder using an XGBoost model

Objective:

To evaluate biomarkers that assist in the identification of electroencephalography signals in patients with depression, utilizing an ensemble learning model.

Methods:

XGBoost ensemble model was trained and tested to classify patients with major depressive disorder (MDD) (n=24) and healthy controls (HCs) (n=29) from the multi-modal open dataset for mental-disorder analysis. Based on event-related potentials (ERPs) to emotional-neutral face pairs (Happy-Neutral, Sad-Neutral, Fear-Neutral) as stimuli, we segmented six conditions: 3 emotional cues, dots (happy, sad, and fear) and applied the FisherScore feature selection method to select the waveform features with high mutual information. Overall, 80% of the data was selected to establish the XGBoost model with five-fold cross-validation.

Results:

We identified happy, sad and fear conditions with waveform features (170–230 s) to distinguish patients with depression. The proposed XGBoost model had a comprehensive accuracy, precision, recall, F1-score, and area under curve of 99.52%, 99.39%, 99.67%, 99.52%, and 99.98% for the ERPs. Furthermore, our experimental results indicated that suppression of the amplitude of negative emotional cues could be used to recognize depression, which was predominantly over the frontal lobe and frontal poles regions. The response latency of ERP signals contributed significantly to distinguishing between HCs and patients with MDD.

Conclusion:

An ensemble learning system for classification using the XGBoost and feature selection using FisherScore has the potential to be used in clinical prediction of depressive symptoms in patients with MDD.

Significance:

The discovery of ERPs as a biomarker has important clinical implications for exploring the pathogenesis behind MDD.