Jin Woo Choi, PhD

Current methods for personalizing adaptive DBS are often subjective and inefficient. The procedures rely on experts manually visualizing neural signals and using trial-and-error to find patterns. This approach uses simplified, low-complexity features and handcrafted parameters, which may fail to capture the detailed neural characteristics needed for optimal, real-time therapy.

This branch introduces deep learning models that aim to automate and personalize this process. Our N2GNet architecture directly translates a user's complex neural signals into continuous gait performance, automatically learning important gait-related neural features through a stepping-in-place task performed on two force plates. To make this technology more accessible, we also developed the N2G Calibrator, which personalizes the model for a new user by leveraging data from others, removing the need for the use of specialized force plates.

While motor imagery is an intuitive paradigm for Brain-Computer Interfaces (BCIs), translating it into real-world device control is challenging. The process requires extensive user training, and controlling a device safely in a dynamic environment often requires additional support apart from noise-prone BCI signals.

This branch involves developing a comprehensive software protocol for user training, simulation, and real-world device control. By integrating BCI within a VR environment and incorporating sensors like LiDAR for shared control, the protocol supports the use of EEG-based motor imagery BCIs for device control.

Classification (decoding) of neural signals is a core component of BCI performance. While current EEG BCIs require extensive data collection from users and rely on expensive, high-channel-count EEG devices to achieve good performance, these demands hinder real-life usage and create a hurdle for mass adoption.

This branch involves advancing deep learning techniques for signal processing. Related components include: developing cross-subject classification models that utilize data from other individuals to aid a new user; enhancing performance on artifact-prone EEG recordings; and enabling integration with EEG devices that have fewer electrode channels.