Evidence for decoder-label-regime-dependent effects of complementary leaky integrate-and-fire temporal-state decoding in motor-imagery EEG

Background Motor-imagery electroencephalography often provides limited subject-specific supervision. This study tested whether complementary leaky integrate-and-fire (CLIF)-family temporal-state decoding shows a decoder-level advantage after the EEG representation is fixed, how that advantage changes with decoder-label availability, and whether hard binary spikes are required. Methods ANN-SEW, the artificial neural network comparator with spike-element-wise (SEW) residual topology, and SEW-CLIF received identical frozen attention-based temporal convolutional network (ATCNet) and filter-bank convolutional network (FBCNet) representations. The evidence chain combined BCI Competition IV Dataset 2a (BCI2a) development, retrospective OpenBMI replication, an internally pre-access-held-out BCI Competition IV Dataset 2b (BCI2b) evaluation, cross-entropy-only label-budget analyses, and a separate internally pre-access-locked BNCI2015-001 Session-A-to-Session-B evaluation. Additional robustness analyses included shared whole-sequence readout for parameter-matched gated recurrent unit (GRU) and temporal convolutional network (TCN) decoders, exact-gradient sigmoid-gated Soft-CLIF, equal optimizer-update budgets, a within-BCI2a binary sensitivity analysis, and computational profiling. Results Under cross-entropy-only training, SEW-CLIF exceeded ANN-SEW by 10.082 percentage points (pp) on BCI2a and 5.470 pp. on OpenBMI, while full-budget BCI2b remained near parity (+0.181 pp). The ANN-relative gap decreased as decoder labels increased and persisted under equal-update training. At 25 labels/class, exact-gradient Soft-CLIF exceeded ANN-SEW by 11.085, 8.479, and 5.809 pp. on BCI2a, BCI2b, and OpenBMI, respectively. With matched whole-sequence readout, hard SEW-CLIF was broadly comparable to GRU-Stat and TCN-Stat on BCI2a and BCI2b and modestly higher on OpenBMI. The internally pre-access-locked BNCI2015-001 evaluation showed the same attenuation, with a slope of −2.319 pp. per label doubling (95% CI, −4.118 to −0.604). Conclusion CLIF-family temporal-state dynamics show a reproducible decoder-label-regime effect under frozen supervised representations. Hard forward discretization is not required for the ANN-relative effect, and matched temporal readout substantially reduces differences between CLIF and other continuous temporal decoders.