FOCUS: A Lightweight Mamba-2 Model for EEG-Based Cognitive Workload Classification

Ioannis Kyprakis; Stefanos Gkikas; Chariklia Chatzaki; Zongyi Jiang; Randy Gomez; Raul Fernandez Rojas

cs.AIOct 23, 2026

FOCUS: A Lightweight Mamba-2 Model for EEG-Based Cognitive Workload Classification

Authors: Ioannis Kyprakis, Stefanos Gkikas, Chariklia Chatzaki, Zongyi Jiang, Randy Gomez, Raul Fernandez Rojas

Organizations: Hellenic Mediterranean University, Heraklion, Greece · Honda Research Institute Japan, Wako City, Japan · University of Canberra, Canberra, Australia · BioSIS (Biosensing & Intelligent Systems) Lab, Centre for Intelligent Computing and Systems, University of Canberra, Canberra, Australia

Abstract

Cognitive workload estimation from physiological signals is increasingly relevant in high-risk environments, where continuous and unobtrusive monitoring can inform adaptive systems and safety interventions. This paper presents FOCUS (Frequency-band Optimized Cognitive-load Ultralight SSM), a compact Mamba-2-based model for EEG-based binary workload classification under constrained data conditions. Rather than learning frequency sensitivity from data, FOCUS encodes preprocessed multi-channel EEG using a fixed FFT decomposition that covers the five canonical EEG bands. The proposed model combines depthwise-separable convolution and a weight-shared encoder-bottleneck Mamba-2 pair, with 0.04 million parameters and 0.11 GFLOPs. Evaluated in a pilot study with 11 participants under leave-one-subject-out cross-validation across three cognitively distinct tasks, FOCUS achieves 74.75% accuracy on the full dataset at 0.08 ms GPU model-forward latency. In the weight-sharing ablation, the shared configuration achieves 4.55 percentage points higher accuracy on the full dataset than the non-shared variant while reducing the SSM parameter budget; however, this difference does not reach statistical significance under a two-sided Wilcoxon signed-rank test ( $p=0.083$ ). Overall, the selected FOCUS configuration provides a favorable trade-off between classification performance and computational cost within this controlled pilot setting.

Figures & tables

Fig. 1 : Overview of the proposed pipeline.

		Configuration				Cost		Performance (%)
Config	Task	$d$	$d_{\text{state}}$	DWSep	Shared	Params (M)	GFLOPs	Accuracy	Precision	F1
A1 (selected)	IQ	64	16	✓	✓	0.04	0.11	$67.42{\scriptstyle\,\pm 6.84}$	$69.33{\scriptstyle\,\pm 7.97}$	$66.77{\scriptstyle\,\pm 7.00}$
	MATH	64	16	✓	✓			$\mathbf{75.76}{\scriptstyle\,\pm 7.91}$	$77.00{\scriptstyle\,\pm 7.89}$	$\mathbf{75.43}{\scriptstyle\,\pm 8.06}$
	GAME	64	16	✓	✓			$76.14{\scriptstyle\,\pm 7.96}$	$\mathbf{78.83}{\scriptstyle\,\pm 7.66}$	$75.43{\scriptstyle\,\pm 8.39}$
	ALL	64	16	✓	✓			$74.75{\scriptstyle\,\pm 10.54}$	$76.29{\scriptstyle\,\pm 10.00}$	$74.13{\scriptstyle\,\pm 11.11}$
A2 (no sharing)	IQ	64	16	✓	$\times$	0.07	0.11	$\mathbf{71.97}{\scriptstyle\,\pm 6.90}$	$\mathbf{73.63}{\scriptstyle\,\pm 6.85}$	$\mathbf{71.40}{\scriptstyle\,\pm 7.16}$

TABLE I : Ablation study over architectural configurations. A1 is selected based on the overall performance–cost trade-off.

Model	Task	Params (M)	GFLOPs	Acc (%)	Prec (%)	F1 (%)
EEGNet [ 30 ]	IQ	0.01	0.02	$65.15\pm\pm 6.5$	$65.90\pm\pm 6.6$	$64.75\pm\pm 6.7$
	MATH			$63.64\pm\pm 13.7$	$63.95\pm\pm 13.9$	$63.32\pm\pm 13.8$
	GAME			$64.39\pm\pm 9.5$	$64.93\pm\pm 9.9$	$64.19\pm\pm 9.4$
	mean ∗			64.47
1BT [ 21 ]	IQ	0.45	0.02	$69.32\pm\pm 7.8$	$71.22\pm\pm 8.6$	$68.55\pm\pm 8.2$
	MATH			$67.05\pm\pm 7.6$	$72.73\pm\pm 7.3$	$64.43\pm\pm 10.2$

TABLE II : Comparison of FOCUS with published baselines and a plain Mamba-2 baseline evaluated in the present study, using the same dataset and subject-independent LOSO protocol.

EEG Decoding

Electroencephalography (EEG) decoding maps brain-signal patterns to mental states, stimuli, speech, or intended actions.

See all 205 papers

One-Block Transformer (1BT) for EEG-Based Cognitive Workload Assessment
Apr 21, 2026Stefanos Gkikas, Christian Arzate Cruz, Thomas Kassiotis +3Efficient Neural Network Inference Transformer
Honda Research Institute Japan · Wako City, Japan · Department of Electronic Engineering +6
Assessing Region-Level EEG Contributions to Cognitive Workload Prediction
May 23, 2026Jacob Wong, Sohan Singh, Prannaya Gupta +3Electroencephalography EEG Decoding
School of InfoComm Technology, Ngee Ann Polytechnic · Engineering Product Development Pillar, Singapore University of Technology and Design · NUS High School of Math and Science
FEMBA on the Edge: Physiologically-Aware Pre-Training, Quantization, and Deployment of a Bidirectional Mamba EEG Foundation Model on an Ultra-low Power Microcontroller
Mar 18, 2026Anna Tegon, Nicholas Lehmann, Yawei Li +3EEG Foundation Models Edge ML
Integrated Systems Laboratory, ETH Zürich, Zürich, Switzerland · DEI, University of Bologna, Bologna, Italy

Apr 21, 2026J/K move · Enter open · S save

One-Block Transformer (1BT) for EEG-Based Cognitive Workload Assessment

Stefanos Gkikas, Christian Arzate Cruz, Thomas Kassiotis, Giorgos Giannakakis+2

Honda Research Institute Japan · Wako City, Japan · Department of Electronic Engineering+6

Efficient Neural Network Inference Transformer EEG Decoding

Open paper AskPDF arXiv

Accurate and continuous estimation of cognitive workload is fundamental to creating adaptive human-machine systems. However, designing architectures that balance representational capacity with computational efficiency has been challenging for practical deployment. This paper introduces 1BT, a One-Block Transformer for compact and efficient EEG-based cognitive workload assessment. The model aggregates multi-channel temporal sequences via a minimal latent bottleneck, using a single cross-attention module followed by lightweight self-attention. A controlled study involving 11 participants performing three cognitively diverse tasks (abstract reasoning, numerical problem-solving, and an interactive video game) was conducted with continuous EEG recordings across two workload levels. Systematic architectural analysis identifies the most compact configuration that preserves high performance, while substantially lowering computational cost. The final model achieves high workload classification performance with under 0.5 million parameters and 0.02 GFLOPs, paving the way for a design direction for real-time cognitive workload monitoring in resource-constrained settings.

Abstract

EEG Decoding

Explore similar work

One-Block Transformer (1BT) for EEG-Based Cognitive Workload Assessment

Assessing Region-Level EEG Contributions to Cognitive Workload Prediction

FEMBA on the Edge: Physiologically-Aware Pre-Training, Quantization, and Deployment of a Bidirectional Mamba EEG Foundation Model on an Ultra-low Power Microcontroller