A unified context-aware AI framework for multiscale temporal learning and brain state prediction

The major challenge in computational neuroscience arises from the coexistence of temporal fluctuations, long-term dependencies, and complex context-aware dynamics when predicting brain states from multiscale signals. The existing models, including recurrent models (GRU), hybrid models (CNN-BiLSTM), Transformer-based architectures (EEG-PatchFormer), and graph-based models (GLCNet) have demonstrated effectiveness on sequential modeling but may face limitations in capturing hierarchical temporal patterns and adaptive memory modelling across different timescales. To address these challenges, this paper proposes a Unified Context-Aware AI (UCAI) framework that integrates multiscale temporal learning and adaptive memory with context-aware temporal representation, including brain prediction within a single coherent architecture. Unlike conventional sequential modelling approaches, the proposed framework represents neural signals with structured context-aware temporal representations. A mathematically formulated context-aware temporal representation module maps temporal features into higher-level representations through context-aware transformations to enhance decision-making and contextual representations. The experimental evaluation using “EBrains” (12 subjects) and “SEED-IV” (15 subjects) datasets under Leave-One-Subject-Out (LOSO) cross-validation protocol. The UCAI achieves significant improvement in terms of classification performance and under LOSO evaluation (p < 0.05). In addition, cross-dataset and distribution-shift evaluations provide preliminary evidence of transferability.