Direct information transfer rate optimisation for SSVEP-based BCI

Cover story for the research paper by Anti Ingel, Ilya Kuzovkin and Raul Vicente
Direct information transfer rate optimisation for SSVEP-based BCI
Journal of Neural Engineering, Volume 16, Number 1

---

Information transfer rate (ITR) is de facto standard measure of performance of a brain-computer interface (BCI) system. ITR is measured in bits per minute and can be used to compare performance of different methods regardless of their internal performance or optimization metric.

Most BCI systems rely on machine learning (ML) to decode user’s neural signals into actions of the target system. Standard machine learning approaches run iterative optimization process that minimizes the error between the prediction of the ML model and the ground truth (such as mean squared error or cross-entropy loss). However, if the ultimate purpose of the system is to yield high ITR, it only makes sense to optimize directly for that metric, instead of using the canonical proxies of performance.

In this work Anti derived a formula for calculating gradient of ITR for a BCI system based on steady-state visually evoked potentials (SSVEP). The results confirm that direct optimization for the final metric yield better results and allows for more informed parameter search.

Preprint in arXiv: https://arxiv.org/abs/1907.10509
Publication in Journal of Neural Engineering: https://iopscience.iop.org/article/10.1088/1741-2552/aae8c7/meta
Implementation on GitHub: https://github.com/antiingel/ITR-optimisation