Beyond Words

Being unheard even when you are saying something is probably one of the worst feelings ever. I first truly understood this last summer while volunteering at a speech and behavioural therapy camp for children with autism. According to the National Institute on Deafness and Other Communication Disorders, roughly 25% to 30% of children with ASD are non-verbal, and many more are minimally verbal. One day at camp, there was a boy stimming excessively. We thought he was happy, but back indoors, he began screaming and clutching his stomach.

The leader assumed he was throwing a tantrum. She ended up raising her voice, trying to get him to follow the rules. Minutes later, we realized the truth: he was in pain and needed the bathroom. My heart sank. He was yelled at simply because we could not see what was happening inside his head.

This inspired me to bridge that communication gap using a Brain-Computer Interface (BCI): a system that creates a direct communication pathway between the brain and a computer. I used a Muse 2 consumer-grade EEG headset (Figure 1) and Python to build a BCI that detects Cognitive Overload (mental exhaustion) in real-time. My goal is to replace "lucky guesses" with a definitive voice. Inspired by how dolphins use "sono-pictorial" frequencies to share mental images, I believe we can ensure no one is trapped behind their own silence.

To develop my solution, I focused on turning raw brain activity into a voice for the misunderstood. I wanted software that could handle the messy signals of the real world.

Researching dolphin communication led me to wonder if humans could share reality through frequencies, introducing me to EEG and brain waves (Figure 2). My original goal was to measure the P300 Response: an involuntary brain spike occurring 300 milliseconds after recognizing a need. Lacking access to $10,000 clinical hardware, I pivoted. I decided that if I could not have the smartest hardware, I would build the smartest software using a consumer-grade Muse 2.

My design process involved a four-stage Python pipeline. First, the Petal-LSL library streamed live data from sensors at TP9, AF7, AF8, and TP10 (Figure 3). Second, I built an Artifact Gate. Since a blink looks like a lightning strike, I programmed a 1400-microvolt threshold to delete noise from muscle movements.

Third, a Fast Fourier Transform (FFT), an algorithm that deconstructs complex signals into individual frequency components, broke down messy voltage waves. This allowed me to isolate rhythms like Theta and Alpha. I measured the Theta/Alpha ratio (internal effort divided by calm) and the Alpha/Beta ratio. By comparing these, I can mathematically see if the brain is overwhelmed. Finally, I tested machine learning models using 500 samples. I chose Support Vector Machine (SVM) over Random Forest because SVM found an optimal boundary, whereas Random Forest overfitted the noisy data (Figure 4).

To solve reliability issues, I developed the Stability Engine. Instead of guessing every second, my code analyzes a five-second window using a majority vote system. An alert only triggers if at least three seconds agree the child is in distress. This boosted operational reliability to over 80 percent.

The results of my project proved that smarter software can bridge the gap left by accessible, low-cost hardware. Initially, my Machine Learning model achieved a classification accuracy of 66 percent (Figure 7). While this was a strong start for a prototype, a 34 percent error rate meant that the system was still too unreliable for real-world use. In a caregiving setting, a "lucky guess" from a computer is just as frustrating as a lucky guess from a human. I knew I needed to reach a higher standard of reliability to truly help the children I met at camp who are often misunderstood.

To improve these results, I analyzed the raw data to see exactly how the model was failing (Figure 5). I discovered that most errors were caused by momentary glitches or artifacts. Because I was using a consumer-grade headset with dry electrodes, simple physical actions like a single eye blink, a swallow, or a quick head movement created massive electrical spikes. I used a Support Vector Machine (SVM) to map these "Neural Territories" and find a cleaner boundary between states (Figure 6). However, these spikes were still being misidentified by the SVM as shifts in emotional state. This led to my main finding: human emotions and cognitive states do not flicker in and out in a single millisecond. Instead, they are sustained experiences that happen over several seconds.

I developed the Stability Engine breakthrough to capitalize on this physiological reality. By applying Binomial Distribution theory, I created a majority vote system where the computer analyzes a five-second window of data before making a decision. Instead of reacting to one noisy second that might be a "false positive" caused by a blink, the system only triggers an alert if at least three out of those five seconds agree that the brain is in a state of cognitive overload. This mathematical filter successfully smoothed out the noise and boosted the operational reliability of my system from 66 percent to over 80 percent.

These results prove that my custom ratios, specifically the Theta/Alpha ratio, are effective indicators of internal effort and stress (Figure 8). By dividing the wave representing internal struggle (Theta) by the wave representing calm (Alpha), I can see a clear numerical increase when a user is overwhelmed. My findings show that we do not need clinical-grade equipment or a ten-thousand-dollar lab to get meaningful results. We can use accessible technology and custom code to ensure that a child's internal symphony is finally seen and understood. This project demonstrates that with the right software architecture, we can turn a simple headset into a life-changing communication bridge.

The importance of these results lies in the shift from guessing to knowing. My project proves we do not need clinical-grade, $10,000 equipment to understand a non-verbal child’s internal state. Initially, my raw accuracy was only 66%, a common limitation of low-cost, unshielded consumer sensors. However, by using smarter software to compensate for accessible hardware, I proved a reliable communication bridge can be built using public tools.

The significance of this work is the Stability Engine, which acts as a mathematical shield. Clinical systems use wet electrodes and heavy shielding to keep data clean. My breakthrough uses Binomial Distribution theory to filter out the noise of the physical world, like blinks and movements, through code rather than expensive hardware. When the reliability jumped to over 80%, it confirmed that hardware limitations are just math problems waiting to be solved.

From my results, I learned that the human brain provides a far more honest account of emotion than outward behaviour. The Theta/Alpha ratio is a stable, statistically significant indicator of stress, even on a $250 device.

Ultimately, these findings mean a child never has to be misunderstood during a meltdown. This project provides a blueprint for a future where caregivers use simple tablets and headsets to hear the true signal of the brain. It is a step toward a world where no mind is trapped behind silence, and every child has a way to be heard.

To expand this, I want to move from tracking moods to identifying specific needs using the P300 Response. This is an involuntary spike in the brain that occurs about 300 milliseconds after you see something you recognize or want.

I would set this up by flashing icons for things like water or the bathroom on a screen. When the child sees the right icon, their brain will trigger that quick spike. Unlike the slow trends I measured before, the software will catch this split-second event to act as a thought-controlled pointer for caregivers.

I want to thank my parents for all their support throughout this entire process, from encouraging my original interest in this project to helping me manage my work and research. I also want to thank the Peel Regional Science Fair for the opportunity to showcase my work and for the feedback that helped me refine my project. Their platform allowed me to take my ideas from a home prototype to a national level, and I’m grateful for the chance to share what I’ve learned.

Thank you so much!

Artificial Intelligence

Google. (2024). Gemini (March version) [Large language model]. https://gemini.google.com

Hardware and Software

Interaxon Inc. (2020). Muse 2: Brain sensing headband [Hardware and mobile application]. https://choosemuse.com

MindMonitor. (2023). Muse Mind Monitor [Mobile application]. https://mind-monitor.com

Scientific Concepts

Luck, S. J. (2014). An introduction to the event-related potential technique. MIT Press. (Source for the P300 Response logic).

Oskoei, M. A., & Hu, H. (2008). Support Vector Machine-based classification of EEG signals for control of upper limb assistive devices. Medical Engineering & Physics, 30(10), 1203-1217. (Source for SVM machine learning logic).

Statistical Methods

Wackerly, D., Mendenhall, W., & Scheaffer, R. L. (2008). Mathematical statistics with applications. Cengage Learning. (Source for the Binomial Distribution and Majority Vote logic used in the Stability Engine).