Inspiration
The inspiration for this project came from the vast number of friends and classmates in university that are experiencing some form of depression. Often times they are not able, or do not have the time to seek out diagnoses and receive the help that they need.
What it does
This project focuses on researching the use of Deep Learning to diagnose depression based on the input of EEG data. Our first prototype involves trying to detect whether or not the patient has depression, and in the future we will expand out to detecting different levels of depression as well as diagnosing schizophrenia.
How we built it
This project involves developing a preprocessor to prepare the input data, as well as transforming it from the time domain to frequency/time in the form of spectrograms. These spectrograms were then fed into a convolutional neural network via a custom generator.
The neural net was built using Tensorflow, consisting of 3 Depthwise-Separable Convolution layers as well as two fully connected layers terminating to a binary classifier. The choice of Depthwise-Separable Conv layers was because this is much less computationally intensive during the forward pass, allowing for extremely quick inference times.
Challenges we ran into
The greatest challenge was trying to train a deep model in the limited amount of time we had. While we were lucky to be able to train it on an RTX 3090, time was still rather limited, and we did not have enough time for a full session of hyperparameter search.
Accomplishments that we're proud of
We are quite proud of the teamwork it took to put together the preprocessor and the model. While the validation accuracy was not as high as we hoped, we were quite happy that we were able to break 60% validation accuracy in the time we had.
What we learned
While we cannot conclude that our poor accuracy was a result of using a convolutional NN on the frequency data, as we are yet to find the optimal hyperparameters, we are willing to try a recurrent model some time in the future.
What's next for EEG Depression Diagnosis
After we develop a model with sufficient validation accuracy, our next goal is to train it to predict different forms of depression, as well as other mental conditions like schizophrenia.
Log in or sign up for Devpost to join the conversation.