<200 Word Description
New evidence shows Neurofeedback training is efficacious (https://mitsar-eeg.com/2020/04/01/adhd_neurofeedback-2/). Some research is even showing it as effective as medication for helping with ADHD symptoms. Another study showed children showing boosted intelligence scores after going through such training. We want to make an app and design inexpensive hardware to help bring it to the open-source community with an "as-simple-as-possible" design, where software can be easily expanded on.
The gameplay in the app is simple. The jetpack animal will fly higher and higher, earning additional points. This gives the player extra incentive to train again and again.
We use two machine learning models - a fully-dense MLP and a sliding-window 2D autoencoder, to determine the scale of how hard someone is focusing.
The headphones contain six electrodes located at optimal locations for the brain-wave activity we are detecting. A tone is also played in the headphones that increases with intensity and pitch as the player focuses harder, quickly helping them learn to train their brain.
Attention-deficit/hyperactivity disorder (ADHD) is one of the most common mental disorders affecting children. ADHD also affects many adults. Symptoms of ADHD include inattention (not being able to keep focus), hyperactivity (excess movement that is not fitting to the setting) and impulsivity (hasty acts that occur in the moment without thought). ADHD patients are often involved in traffic accidents, unemployment, divorce, face less possibilities of professional advancement and experiment chronic frustration and underachievement.
EEG biofeedback (neurofeedback) originated in the late 1960s is a method for retraining brainwave patterns through operant conditioning. Since that time a sizable body of research has accumulated on the effectiveness of neurofeedback in the treatment of uncontrolled epilepsy, ADD/ADHD, anxiety, alcoholism, posttraumatic stress disorder, and mild head injuries. Researches confirm the notion that neurofeedback is a clinically efficacious module in the treatment of children with ADHD.
Gevensleben, H., Holl, B., Albrecht, B., Schlamp, D., Kratz, O., Studer, P., ... & Heinrich, H. (2010).
Hammond, D. C. (2007). What is neurofeedback?. Journal of neurotherapy, 10(4), 25-36.
Lofthouse, N., Arnold, L. E., Hersch, S., Hurt, E., & DeBeus, R. (2012). A review of neurofeedback treatment for pediatric ADHD. Journal of attention disorders, 16(5), 351-372.
Neurofeedback training in children with ADHD: 6-month follow-up of a randomised controlled trial. European child & adolescent psychiatry, 19(9), 715-724.
Souza, I. D., Mattos, P., Pina, C., & Fortes, D. (2008). ADHD: The impact when not diagnosed. Jornal Brasileiro de Psiquiatria, 57(2), 139-141.
What it does
Neural Radar provides neurofeedback therapy for ADHD patients, allowing self-regulation of brain function.
Based on the patient's EEG, Neural Radar reads the patient's brain activity and adjusts training for self-regulatory neuromodulation. In this sense, patients can use the headset-like device and control it using a mobile app. Therefore, our goal is to provide an affordable and accessible treatment that can be done at home.
- Start session;
- Generate report;
How we built it
We researched an ideal design by scouring through data from the internet and scientific journals. We found exactly where electrodes should be placed (image of graph of locations vs efficacy) and made sure there was strong evidence before we set out to make this device for everyone. We got our dataset from IEEE access port (https://ieee-dataport.org/open-access/eeg-data-adhd-control-children).
In a nutshell, the electrodes record how strongly you are focusing, and you receive feedback. In this case, it is a game - the customizable jetpack character will soar higher as you are able to focus stronger. This teaches you how to focus better, and similar techniques have been shown to significantly improve ADHD symptoms, with even comparable performance to medications. One study found that even non-ADHD children benefited, and were able to perform higher on IQ tests.
Software and services we used in this project were Python, Flutter, Blender, PyTorch, Tensorflow, Heroku and AWS.
The simplest way to build this is with a TI ADS1299, a ~$35 chip that is an 8-channel EEG. It can interface with an Arduino or Pi that we would connect to a phone via Bluetooth, allowing a user to use our app.
Challenges we ran into
Learning Flutter as we went along. One of the major challenges we ran into is Tuning MLP to classify electrodes data efficiently. The electrodes data for control childrens (childrens who are not diagnosed with ADHD) and ADHD childrens had very little distinguishable feature at time. This would throw off the learning rate for the Machine learning model. On the first try, we hit a saddle point most likely due to barely distinguishable data points at times. One the second try, our learning rate was all over the place due to very low value for learning rate and small batch sizes. We had to play around with the parameters to get the desired result.
Accomplishments that we're proud of
Everyone has their name on a significant segment of the project and we had a great time collaborating in the white-board room! Being able to combine significant contributions from all team members was the most important accomplishment for us. We initially set out to work collaboratively on the project, and we were able to meet this goal!
One of the machine learning model we leveraged was a good old back propagating Multilayer Perceptron. Since we had only 6 input layers to focus on, our MLP only input 6 layers for processing. The results were surprisingly accurate. In such a short amount of time, we were able to reach upto 80% training accuracy. We believe with more time, and more refined tuning, we can get the accuracy even higher. We deployed the MLP classifier to Heroku for processing the data over the internet using Flask.
A sliding-window autoencoder runs on-device via tflite. Three EEG readings are transformed into two-dimensional time-series data, and the autoencoder uses low-density convolutional, deconvolutional, and fully dense layers with a minimal latent space to measure reconstruction loss to determine how far off the user is from their potential peak thinking. A LAMB with an exponential cyclical learning rate was used to train both the encoder and decoder portions. Remarkably clear results were found between ADHD and the non-ADHD controls.
What we learned
One of the major aspect we learnt is that determining ADHD and focus span is not as black and white as people may believe. It is not something that can be classified as a binary identity, and has a lot of other factors in play. The more we analysed the data, the more we realize how intricate is the world of neural feedback.
Bill of materials
What's next for Neural Radar
Currently we demonstrate the potential of combining Mobile, AI and hardware for cost efficient neurofeedback. In the future, we will continue to develop the project, building and improving the hardware model as well as the software behind it to contribute to the opensource community. We would like to integrate Database, in the future, to save patients reports and information for further processing and analysis.