Diagnosing Motor-Imagery BCI Failure: Separating Methodological from Neural Decoding Failure

Many people cannot control a motor-imagery brain–computer interface (BCI) well. This problem is often called BCI illiteracy. Most research tries to fix it by using better decoding methods, assuming that everyone can eventually succeed. That assumption is rarely tested. In this study, we propose a way to tell apart two different reasons for failure: a method failure, where the brain signal has useful information but the decoding method misses it, and a neural failure, where the signal contains too little discriminable structure for usable control, and even stronger methods cannot extract more. We used a public dataset with nine subjects performing four motor-imagery tasks. We created a separability score based on the geometry of the EEG signal, without training any classifier. This score predicted how well a simple decoder would work (r = .89). The three worst subjects all got low separability scores and clustered together. We then tried two stronger decoding methods: filter-bank CSP (FBCSP) and a Riemannian minimum-distance classifier (MDM). One subject (S5) improved a lot with FBCSP, going from a kappa of .28 to .55. The other two (S2 and S6) obtained similarly low absolute kappa under all three decoding pipelines, although S6's gain over the baseline became statistically significant when eye artifacts were removed. This shift moved S6 from the neural-failure group to the method-failure group. It shows that the diagnosis depends on how the signal is prepared. Our findings indicate that BCI illiteracy mixes up different kinds of problems. A two-step approach, first a classifier-free screen and then a test of whether a better method helps, can sort out which subjects are worth trying to rescue and which are not, while being honest about the limits of the tools.