A Multi-Scale Temporal Framework with Dynamic Fusion for EEG-Based Emotion Recognition

This study introduces a multi-scale temporal framework for EEG-based emotion recognition. The EEG waveform is decomposed into windows of one or several durations, processed by a shared attention-based encoder, and integrated through a dynamic fusion module that assigns sample-specific weights across temporal scales. The framework is evaluated under a subject-independent protocol in binary and three-class settings, with the three-class task including the mixed affective category. The best results are 65.22% for the two-class task and 45.43% for the three-class task. Both are obtained with three-scale dynamic-fusion configurations and remain substantially above the full-signal baseline. The best-performing temporal scales differ between the two tasks. Dynamic fusion outperforms concatenation in the highest-scoring two-class configuration and slightly exceeds it in the highest-scoring three-class configuration, although these multi-scale settings require substantially more computation than the full-signal baseline.

Improvements to AI Systems:

1. Adaptive Multi-Scale Temporal Encoding: Implement a shared attention-based encoder that processes EEG signals at multiple, dynamically-selected window durations (e.g., 1s, 3s, 5s) in parallel. This allows the AI to capture both fast transient neural patterns and slower sustained emotional states simultaneously, improving robustness to temporal variability in affective responses.

2. Sample-Specific Dynamic Fusion: Replace static feature concatenation with a learned dynamic fusion module that assigns per-sample, per-scale attention weights. This enables the AI to automatically emphasize the most informative temporal scale for each individual trial, adapting to inter-subject differences in emotional expression latency and duration.

3. Task-Adaptive Scale Selection: Use a meta-learning or hyperparameter optimization layer to automatically select the optimal set of temporal scales for a given classification task (e.g., binary vs. three-class with mixed affect). This prevents manual tuning and improves generalization across different emotional taxonomies.

4. Computational Efficiency via Sparse Fusion: Introduce a sparsity or pruning mechanism in the dynamic fusion module to reduce the computational overhead of multi-scale processing. This allows the improved system to run on edge devices (e.g., wearable EEG headsets) while retaining the accuracy gains from multi-scale analysis.

What the Improved AI System Can Do:

• Real-time Emotion Tracking: Continuously monitor EEG from a wearable device and classify emotional states (e.g., positive/negative, or valence/arousal/mixed) with higher accuracy than single-scale baselines, adapting to each user’s unique neural timing.

• Personalized Affective Computing: Automatically recalibrate temporal attention weights per session or per user, improving performance in brain-computer interfaces for mental health monitoring, adaptive learning, or gaming.

• Robust Cross-Subject Deployment: Maintain high classification accuracy in subject-independent scenarios (e.g., new users with no prior training data) by leveraging multi-scale features that are less prone to overfitting to individual temporal patterns.

• Task-Specific Optimization: Switch between binary and fine-grained three-class emotion recognition without manual feature engineering, using the dynamic fusion to emphasize the most discriminative temporal scales for each task.

• Energy-Aware Operation: Optionally run in a low-computation mode (fewer scales, sparse fusion) for battery-constrained devices, while retaining a high-accuracy mode for offline analysis or clinical use.