In this project, I aimed to predict Major Depressive Disorder (MDD) based on an open dataset of electroencephalograms (EEGs). I not only developed a basic Convolutional Neural Network (CNN) model for classification but also explored advanced techniques, including swarm learning with HPE (Hewlett Packard Enterprise) and generating new EEGs using a Diffusion model.
In this chapter, I performed essential data preprocessing tasks to prepare the EEG data for analysis. The preprocessing involved:
- Tuning windowing parameters to achieve optimal segmentation of EEG signals for upsampling.
- Removing unnecessary channels.
- Applying differnt filters (a notch filter, high and lowband filter, artifact removal with mne-library).
- Filter for channels that could be potentially labeled wrongly with trusted learning with cleanlab
In this chapter, I created a basic CNN model for MDD classification using the preprocessed EEG data. The model achieved an accuracy of up to 91 percent when coupled with the right preprocessing techniques.
This chapter focuses on the distribution of data among peers in a swarm learning solution provided by HPE. Unfortunately, due to confidentiality agreements from my internship at HPE, I can't share specific implementation details or outputs. However, I can confirm that the results were comparable to those achieved by the local algorithm.
In this chapter, I developed my own diffusion model tailored to work with 2D EEG data. The model performed well, successfully reproducing EEG signals with some notable observations:
- Specific noising step for EEG data
- Overfitting on a single EEG file yields the same EEG with a bit of noise which proves the functionallty of my U-Net
- Some runs which specific EEGs generated EEGs with similar frequency bands but higher noise levels.
- An attempt to use a 1D Unet model yielded mixed channel information, which requires further investigation.
Please note that certain details and results may be limited or confidential due to the nature of the project.