Over the last few years, the use of deep learning models in EEG-based depression detection has shown significant progress, achieving high accuracy in the classification between healthy and depressed subjects. However, most of the existing studies did not consider crucial issues affecting performance: How demographic factors like gender can bring biases in EEG signal characteristics, the impact of task variability on the diagnosis accuracy. Traditional approaches have strongly relied on resting-state EEG data, but incorporation of task-based recordings can be used to further understand the cognitive states associated with depression. Moreover, this transparency of the model can be enhanced through the use of XAI techniques, which provides clinicians insight into how decisions are being made and raises confidence in the predictions. This reviews also explore the potential of combining deep learning and XAI might produce more interpretable models for clinical application.