Abstract
<title>Abstract</title> <p>Brain–computer interfaces built for neurorehabilitation rarely get to work with complete information. A stroke patient, or someone with spinal-cord damage, may only be able to express a broad intention, “move this limb,” without the fine-grained motor detail a classic decoder expects. We compared a hierarchical cascade classifier against a flat multiclass classifier on the same four-class motor imagery task, drawing on two public datasets, EEGMMIDB with n=45 subjects and BCI Competition IV Dataset 2a with n=9 subjects, both reduced to covariance-based tangent-space features and logistic regression. In balanced accuracy the cascade shows no statistically significant difference from the flat model, 0.501 versus up to 0.517 on EEGMMIDB and 0.622 versus 0.616 on IV–2a, yet it needs 25% fewer parameters and lets the system activate decision levels progressively, following the patient’s residual capability. What the comparison shows is that the useful information is naturally arranged into difficulty levels. The coarse limb-type decision already holds most of the usable signal, with normalized skill climbing from 0.20 at the first depth to 0.45 at the second, while the left/right lateralization decision is the consistent bottleneck in both datasets, AUC 0.66 to 0.78, a finding backed by contralateral sensorimotor topographies. A selectable reject option adds a small gain in the coverage–accuracy trade-off for the cascade. For adaptive neurorehabilitation, we argue, the clinically relevant property is not a marginal gain in accuracy but the capacity to decompose activity into actionable levels and to keep working under partial information.</p>