Skip to main content
. 2025 Aug 11;16:1596408. doi: 10.3389/fneur.2025.1596408

Figure 4.

Diagram of the Attention-Guided Optimization Strategy (AGOS) featuring three main components. The Attention-Driven Feature Selection uses Gabor, mixing, convolutional, and FC layers for EEG and EOG data. Domain-Constrained Optimization employs LSTM layers. The Temporal-Aware Attention Mechanism applies normalization, dropout, batch normalization, ReLU, and reshaping.

The architecture of the Attention-Guided Optimization Strategy (AGOS), integrating Attention-Driven Feature Selection, Domain-Constrained Optimization, and a Temporal-Aware Attention Mechanism. AGOS employs convolutional layers, LSTM networks, and attention mechanisms to enhance feature selection, align predictions with clinical knowledge, and capture temporal dependencies in medical time-series data. The Attention-Driven Feature Selection module extracts meaningful information using Gabor layers, convolutional processing, and fully connected layers, ensuring that clinically relevant features are emphasized while suppressing noise. The Domain-Constrained Optimization framework enforces fairness, clinical alignment, and temporal consistency by integrating domain-specific constraints into the loss function, improving both interpretability and robustness. The Temporal-Aware Attention Mechanism dynamically adjusts attention weights over time, capturing critical phases in disease progression and allowing the model to focus on pivotal moments in patient trajectories.