Adaptive local temporal weighting for short-window EEG-based SSVEP recognition using FB-eTRCA.

To improve short-window SSVEP recognition, we integrated adaptive temporal local weighting (ATLW) with filter-bank ensemble task-related component analysis (FB-eTRCA). ATLW estimates local discriminability from training trials, converts window scores into sample weights, and applies the training-derived weights to training and test trials within each cross-validation fold. Experiments were conducted on two public 40-target SSVEP datasets: the Benchmark dataset containing 35 subjects and six blocks, and the BETA database containing 70 subjects and four blocks. Seven observation windows ranging from 0.2 to 1.4 s were evaluated using subject-specific leave-one-block-out validation, with six folds for the Benchmark dataset and four folds for the BETA database. On the Benchmark dataset, only the 0.2-s accuracy gain remained significant after Holm correction (44.79% vs 44.07%; adjusted p = 0.0300); other windows were nonsignificant. On the Benchmark dataset, ablation analysis showed that adaptive weighting was more stable than fixed temporal weighting. After Holm correction, eTRCA+ATLW significantly outperformed eTRCA at 0.4 s (gain: 0.86 percentage points; adjusted p = 0.0023), but not at 0.2 s, supporting transferability to another TRCA-based decoder. EEGNet and BiLSTM were included as representative CNN and RNN baselines; both performed below the TRCA-family methods under the present subject-specific limited-calibration protocol. On the BETA database, FB-eTRCA+ATLW remained close to FB-eTRCA across all seven observation windows without BETA-specific parameter retuning, but no paired difference was statistically significant. These findings indicate that ATLW is a lightweight temporal refinement that can be applied to more than one TRCA-based decoder. Its measurable benefit may depend on SSVEP signal quality and the amount of available subject-specific calibration data. Accuracy increased with longer windows, whereas the information transfer rate peaked at 0.6 s on the Benchmark dataset. Therefore, no observation window was optimal for both accuracy and communication speed.