Time Window Optimization for Motor Imagery Classification via Temporal Structure Embedded Reinforcement Learning

Huan Luo; Na Lu; Xiaopeng Wang; Ziyang Shang

cs.NEOct 12, 2026

Time Window Optimization for Motor Imagery Classification via Temporal Structure Embedded Reinforcement Learning

Authors: Huan Luo, Na Lu, Xiaopeng Wang, Ziyang Shang

Organizations: School of Future Technology, Xi’an Jiaotong University, Xi’an, 710049, China · School of Automation Science and Engineering, Xi’an Jiaotong University, Xi’an, 710049, China

Abstract

Online motor imagery-based brain-computer interfaces (MI-BCIs) typically decode continuous EEG using fixed-length sliding windows that may be misaligned with neural events. Existing window optimization methods often evaluate windows independently, overlooking temporal dependencies and potentially producing fragmented or physiologically implausible selections. We propose Temporal Structure Embedded Reinforcement Learning (TSE-RL), a framework that incorporates prior knowledge of MI event structure into a constrained Markov decision process for adaptive window optimization. First, TSE-RL assigns each window an explicit semantic role through Discard, Onset, Sustain, or Offset actions, extending binary retain/discard decisions to event-level alignment. Second, a finite-state machine masks invalid actions to enforce the temporal order of MI phases and promote physiologically plausible window sequences. Third, a Transformer-based history encoder captures dependencies across windows, improving training stability and computational efficiency compared with recurrent architectures. TSE-RL is plug-and-play and can be integrated with diverse MI classifiers. Experiments on multiple benchmark datasets demonstrate consistent improvements in decoding performance in both cross-session and cross-subject settings. Visualization analyses further suggest that the learned window selection policy aligns with neurophysiological principles.

Figures & tables

Fig. 1: Overview of the proposed TSE-RL framework. (a) MDP formulation with an internal FSM for action masking. At step $n$ , the agent receives the state $s_{n}$ from the environment. The FSM determines the valid action subset $\mathcal{A}_{\phi_{n}}$ for the current phase $\phi_{n}$ and constrains the masked policy $\tilde{\pi}_{n}$ . The selected action $a_{n}$ drives the environment transition and the reward. (b) Network architecture of TSE-RL. The feature extractor encodes each window into a representation, and the discriminator outputs class probabilities for reward computation. A Transformer-based history encoder models temporal dependencies among historical features. The FSM-constrained policy network integrates current features, historical context, and trial progress to produce actions under phase-dependent masking.

Dataset	Number of subjects	Number of channels	Sampling rate (Hz)	Trial length (s)	Number of sessions	Trials in 1st session	Types of imaginations
BNCI2014001	9	22	250	4	2	288	r/l hand, feet, tongue
BNCI2014002	14	15	512	5	1	160	right hand, both feet
BNCI2015001	12	13	512	5	3	400	right hand, both feet
OpenBMI	54	62	1000	4	2	200	right/left hand

TABLE I: Summary of the four MI-EEG datasets

Setting	Method	S0	S1	S2	S3	S4	S5	S6	S7	S8	Avg.
Baseline-full	EEGNet	84.97	53.51	86.15	61.92	54.14	52.54	76.85	80.11	81.16	$70.15_{\pm 2.04}$
	TSLANet	88.38	51.22	84.56	47.40	37.68	58.86	77.61	84.56	82.12	$\mathbf{68.04}_{\pm 2.43}$
	EEGConformer	84.51	49.83	89.19	63.41	46.57	55.86	86.50	79.30	81.73	$70.77_{\pm 0.92}$
	FBCNet	82.33	58.38	81.64	68.79	44.14	51.78	84.07	83.38	84.07	$70.95_{\pm 1.58}$
	ATCNet	87.26	54.56	89.21	65.95	57.26	59.07	82.96	83.72	83.65	$73.74_{\pm 1.33}$
	M-ShallowConvNet	86.50	54.84	91.50	70.94	57.47	58.86	89.07	83.51	83.65	$75.15_{\pm 1.21}$

TABLE II: Within-subject classification accuracy (%) in simulated online experiment on BNCI2014001, with an asterisk () denoting the significance level (: p¡0.05).

Setting	Method	BNCI2014002	BNCI2015001	OpenBMI
Baseline-full	EEGNet	$85.59_{\pm 1.91}$	$82.42_{\pm 1.96}$	$71.97_{\pm 1.30}$
	TSLANet	$83.33_{\pm 2.80}$	$85.79_{\pm 2.64}$	$73.67_{\pm 1.65}$
	EEGConformer	$76.83_{\pm 1.54}$	$85.29_{\pm 1.67}$	$73.94_{\pm 1.31}$
	FBCNet	$87.26_{\pm 1.56}$	$83.91_{\pm 1.53}$	$74.93_{\pm 1.09}$
	ATCNet	$88.81_{\pm 1.25}$	$87.96_{\pm 1.29}$	$75.46_{\pm 1.35}$
	M-ShallowConvNet	$88.73_{\pm 1.27}$	$86.37_{\pm 1.14}$	$74.67_{\pm 1.48}$

TABLE III: Average within-subject classification accuracy (%) in simulated online experiments on BNCI2014002, BNCI2015001 and OpenBMI . * indicates

p<0.05

Setting	Method	S0	S1	S2	S3	S4	S5	S6	S7	S8	Avg.
Baseline-full	EEGNet	67.40	31.99	66.01	39.28	30.60	36.50	36.85	69.49	48.31	$47.38_{\pm 2.51}$
	TSLANet	64.97	29.21	62.89	38.24	26.43	33.72	38.93	63.58	55.25	$\mathbf{45.91}_{\pm 2.72}$
	EEGConformer	68.79	33.72	67.06	44.14	28.86	34.42	44.14	70.53	51.08	$49.19_{\pm 1.33}$
	FBCNet	63.24	32.33	80.60	40.67	31.99	33.03	53.17	76.08	47.26	$50.93_{\pm 1.65}$
	ATCNet	70.53	32.33	70.53	46.57	28.86	34.07	42.75	75.04	59.76	$51.16_{\pm 1.44}$
	M-ShallowConvNet	71.22	33.72	69.49	39.28	33.03	32.68	39.28	73.31	53.17	$49.46_{\pm 1.47}$

TABLE IV: Cross-Subject Classification Accuracy (%) in Simulated Online Experiment on BNCI2014001. * indicates

p<0.05

Fig. 2: Average decision process for Left-hand MI on BNCI2014001. (a) Induced ERD/ERS maps computed following Graimann et al. [ 35 ] with bootstrap significance testing (black contours indicate

\alpha<0.05

). (b) Action distribution over time shown as a stacked profile ordered from bottom to top as Discard, Onset, Sustain, and Offset. (c) Mean classification confidence for the true class at each time window.

Fig. 3: Average decision process for Right-hand MI of subject 0. Layout and analysis pipeline are identical to Fig. 2 .

Setting	Backbone	Time (ms)	FLOPs (M)	Params (M)
Baseline (FE+Disc)	EEGNet	0.57	3.13	0.002
	TSLANet	2.73	7.94	0.366
	EEGConformer	4.16	14.39	0.278
	FBCNet	0.42	0.40	0.002
	ATCNet	4.00	12.29	0.036
	M-ShallowConvNet	0.49	16.37	0.014

TABLE V: Single-window computational efficiency under Baseline and TSE-RL settings. Time is measured in ms per window (batch size 1). FLOPs and Params are reported per forward pass.

Components			BNCI2014001		BNCI2014002	BNCI2015001
RL	Action mask	History enc.	within-subject	cross-subject	within-subject	within-subject
$\times$	$\times$	$\times$	$70.15_{\pm 2.04}$	$47.38_{\pm 2.51}$	$85.59_{\pm 1.91}$	$82.42_{\pm 1.96}$
$\checkmark$	$\times$	$\times$	$69.10_{\pm 2.11}$	$48.07_{\pm 2.58}$	$86.54_{\pm 1.81}$	$83.32_{\pm 1.80}$
$\checkmark$	$\checkmark$	$\times$	$70.75_{\pm 1.82}$	$48.72_{\pm 2.36}$	$86.86_{\pm 1.78}$	$83.87_{\pm 1.71}$
$\checkmark$	$\times$	$\checkmark$	$71.41_{\pm 1.86}$	$49.48_{\pm 2.17}$	$87.02_{\pm 1.81}$	$84.25_{\pm 1.87}$
$\checkmark$	$\checkmark$	$\checkmark$	$\mathbf{72.34_{\pm 1.91}}$	$\mathbf{50.15_{\pm 2.25}}$	$\mathbf{87.62_{\pm 1.67}}$	$\mathbf{84.88_{\pm 1.71}}$

TABLE VI: Ablation study on the EEGNet backbone across settings and MI datasets. Mean

\pm

s.d. classification accuracy (%).

Fig. 4: Sensitivity analysis of window length

L

, stride step

\Delta

, and history length

k

on BNCI2014001 (within-subject), BNCI2014001 (cross-subject), BNCI2014002 (within-subject), and BNCI2015001 (within-subject). The left column displays classification accuracy, while the right column shows the ratio of retained windows.

MI Classification

Motor imagery classification decodes imagined movements from EEG to enable brain-computer interface control.

See all 33 papers

ATCNet-CIAM for Multi-Session Motor Imagery EEG Signal Classification
Jul 26, 2026Le Huu Son Hai, Nguyen Chi Hai, Truong Viet Vu +3MI Classification Brain-Computer Interfaces
Faculty of Information Technology, Van Lang School of Technology, Van Lang University, Ho Chi Minh City 70000, Vietnam
Multi-Feature Riemannian Hypergraph for Online Test-Time Adaptation of Motor Imagery Brain-Computer Interface
Aug 17, 2026Siqi Li, Zhi Li, Tong Liu +5Cross-Session EEG Decoding MI Classification
Peking University · Chinese Institute for Brain Research, Beijing · NeuCyber Neurotech +2
EEG-MFTNet: An Enhanced EEGNet Architecture with Multi-Scale Temporal Convolutions and Transformer Fusion for Cross-Session Motor Imagery Decoding
Apr 7, 2026Panagiotis Andrikopoulos, Siamak MehrkanoonMI Classification Brain-Computer Interfaces
Department of Information and Computing Sciences, Utrecht University, Utrecht, Netherlands

Jul 26, 2026J/K move · Enter open · S save

ATCNet-CIAM for Multi-Session Motor Imagery EEG Signal Classification

Le Huu Son Hai, Nguyen Chi Hai, Truong Viet Vu, Nguyen Phuc Nguyen+2

Faculty of Information Technology, Van Lang School of Technology, Van Lang University, Ho Chi Minh City 70000, Vietnam

MI Classification Brain-Computer Interfaces EEG Decoding Temporal Convolutional Networks Cross-Session EEG Decoding

Open paper AskPDF arXiv

Motor imagery (MI)-based electroencephalography is widely used in non-invasive brain--computer interfaces (BCIs), but robust decoding remains challenging due to inter-subject variability and cross-session non-stationarity. This work proposes ATCNet-CIAM, an enhanced attention temporal convolutional network that integrates a lightweight channel-integrated attention module (CIAM) into the ATCNet framework to improve channel-spatial feature representation for MI decoding. The proposed model is evaluated on BCI Competition IV-2a, BCI Competition IV-2b, and the multi-day WBCIC-MI dataset under standard, within-session, and cross-session protocols. Experimental results show that ATCNet-CIAM achieves 86.32% accuracy on BCI IV-2a and 87.96% on BCI IV-2b under the standard protocol, while reaching 89.46% and 83.64% in the within-session WBCIC-MI on 2C and 3C, respectively. The proposed framework consistently improves classification stability and robustness under session-varying conditions, and ablation study confirms the complementary contribution of the proposed architectural components.

Abstract

MI Classification

Explore similar work

ATCNet-CIAM for Multi-Session Motor Imagery EEG Signal Classification

Multi-Feature Riemannian Hypergraph for Online Test-Time Adaptation of Motor Imagery Brain-Computer Interface

EEG-MFTNet: An Enhanced EEGNet Architecture with Multi-Scale Temporal Convolutions and Transformer Fusion for Cross-Session Motor Imagery Decoding