Hybrid CNN–LSTM network for cross-subject EEG emotion recognition

Hashim Ali, Zangar Ermaganbet, Muhammad Tahir Akhtar

2026

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.1007/s44163-026-01314-z
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Detecting human emotions from brain signals is fundamental for building intuitive and adaptive human–computer interfaces (HCI); However, Electroencephalography (EEG) remains challenging to interpret because the signals are non-stationary, high-dimensional, and noisy. This study evaluated whether an attention-guided hybrid deep-learning architecture can deliver reliable, subject-independent emotion recognition from EEG. To answer this, a lightweight CNN–LSTM–CBAM network model is proposed that (i) extracts spatial–frequency features from time–frequency spectrograms with a convolutional neural network (CNN), (ii) captures temporal dynamics in principal-component EEG time-series with a long short-term memory (LSTM) layer, and (iii) refines both via a convolutional block attention module (CBAM) that applies channel- and spatial-wise attention. The model has been evaluated on the four-class SEED-IV dataset (happy, sad, fear, neutral) using a rigorous leave-one-subject-out (LOSO) protocol across 15 participants. Compared with a spectrogram-only CNN (64.21% accuracy) and a CNN–LSTM ensemble lacking attention (64.63%), the proposed network achieves 68.67% cross-subject accuracy and a 68.22% macro-F1 score, while substantially reducing class imbalance and raising the accuracy of the most challenging class, ‘happy’, by 5.4% points. A Wilcoxon signed-rank test on per-subject accuracies confirmed that the 4.0%-point improvement over the baseline, CNN–LSTM ensemble model without attention, has been statistically significant ( \(p<0.01\) ). In contrast to recent literature (2020–present), including recent online-first publications, this method offers competitive performance with significantly fewer parameters and consistent subject-to-subject results. These findings demonstrate that channel-and spatial attention embedded in a hybrid CNN–LSTM pipeline can mitigate EEG noise, balance class performance, and advance practical, real-time affective brain–computer interfaces.