2025Fall BME LAB Final Project Group 5 Brain Computer Interface Game Control System

As Brain-Computer Interfaces (BCIs) evolve from clinical tools to immersive Human-Computer Interaction modalities, the field is shifting from discrete commands toward continuous mental state decoding. This study investigates the feasibility of utilizing continuous EEG fluctuations to control a physics-based simulation governed by stability constraints. We developed "Tightrope Balance," a real-time bio-feedback application driven by a custom Convolutional Transformer Network (CTNet). Trained on a 35-subject dataset to differentiate "Focus" and "Relax" states, the model operates within a decoupled Client-Server architecture supporting both laboratory-grade (BIOPAC) and consumer-grade (BrainLink) hardware. The system maps inference outputs to rotational torque vectors in a Pygame physics engine, creating a continuous control loop, while repurposing ocular artifacts as discrete jump triggers. Rigorous Leave-One-Subject-Out (LOSO) cross-validation yielded a mean accuracy of 71.3%, significantly surpassing the 55.1% baseline and confirming robust subject-independent generalization. In real-time deployment, the system successfully established a responsive bio-feedback loop, allowing users to actively modulate avatar equilibrium. While the custom classifier demonstrated exceptional stability on high-fidelity hardware, cross-platform deployment to wearable devices revealed expected performance trade-offs due to domain shift, though functional control was maintained. Ultimately, this project validates the efficacy of hybrid deep learning architectures in decoding non-stationary EEG signals for continuous control tasks, offering an intuitive, gamified interface for users to master cognitive stability.