MTDANet: lightweight multi-scale temporal hierarchy and distance-regularized linear attention network for motor imagery brain–computer interfaces

Abstract Objective . Motor-imagery (MI) electroencephalography (EEG) offers a non-invasive route to brain–computer interaction, yet reliable single-trial decoding remains difficult because task-related patterns are weak, nonstationary, and distributed over multiple temporal scales. This study investigates whether a staged local-to-global model can integrate fine temporal–spatial structure with longer-range context without relying on unrestricted self-attention. Approach. We propose MTDANet, an architecture that combines multi-scale temporal–spatial feature extraction, factorized temporal consolidation, and distance-regularized recurrent attention. The model was evaluated on three public MI datasets under subject-specific and leave-one-subject-out (LOSO) protocols. Separately initialized models were also trained on P300 and steady-state visual evoked potential (SSVEP) datasets to examine broader architectural applicability. Participant-level paired analyses used two-sided paired t tests with within-setting Holm correction. Main results. MTDANet achieved subject-specific accuracies of 85.61 ± 8.32 % , 90.85 ± 7.62 % , and 87.79 ± 13.93 % on brain–computer interface (BCI) Competition IV 2a, BCI Competition IV 2b, and BNCI2014_002, respectively. Relative to the highest-mean baseline, the paired improvements were supported after within-setting Holm correction on BCI Competition IV 2a (3.56 percentage points, adjusted p = 0.0097 ) and 2b (1.62 points, adjusted p = 0.0168 ). Under LOSO evaluation, the corresponding accuracies were 63.31 ± 15.74 % , 79.77 ± 7.57 % , and 79.96 ± 11.95 % . The 2.08-point BNCI2014_002 difference had a 95% confidence interval of 0.42–3.73 points and an adjusted p value of 0.0387. These three comparisons did not remain significant in a broader 56-comparison Holm sensitivity analysis. The P300 and SSVEP models attained the highest observed mean accuracies. Significance. The results provide within-setting evidence for staged local-to-global temporal modeling in offline MI-EEG evaluations.