Abstract
Accurately measuring cognitive load in real-time is essential for safety-critical applications like driver assistance and adaptive learning. However, processing multimodal physiological signals (EEG, ECG, EDA) typically requires complex, hand-crafted feature extraction. We present UniPhyNet, an end-to-end deep learning architecture that learns representations directly from raw physiological data. By integrating multiscale parallel convolutions, attention-enhanced ResNet blocks, and bidirectional GRUs, UniPhyNet achieves state-of-the-art performance on the CL-Drive dataset, reaching 80.16% accuracy in binary cognitive load classification.
Introduction
Monitoring cognitive load provides objective insights into a user's mental state, which is vital for human-computer interaction. Despite the richness of physiological data, fusing signals with disparate sampling rates and characteristics remains a significant challenge. Traditional methods rely on unimodal data or labor-intensive manual feature engineering, which can miss subtle cross-modal patterns. UniPhyNet addresses these limitations by providing a unified framework for multimodal raw signal classification, bypassing the feature engineering bottleneck entirely.
Methods
UniPhyNet processes raw, preprocessed 10-second signal windows through a multi-stage pipeline designed to capture both local and global temporal features.
The architecture consists of three primary components:
- Parallel Convolutional Blocks: To mimic hierarchical sensory processing, the model employs multiple 1D convolutions with varying kernel sizes (e.g., 3 and 9). This allows the simultaneous capture of high-frequency oscillations and long-term slow-wave patterns.
- ResNet Blocks with CBAM: Standard convolutions are replaced with ResNet blocks integrated with a Convolutional Block Attention Module (CBAM). This spatial and channel attention mechanism enables the network to focus on the most informative parts of the signal while suppressing noise.
- Bidirectional GRU (Bi-GRU): A Bi-GRU layer is used to capture long-term temporal dependencies across the input windows.
For multimodal configurations, UniPhyNet utilizes intermediate fusion. Feature maps from individual EEG, ECG, and EDA branches are concatenated and refined through an attention mechanism to prioritize the most relevant cross-modal features.
Experimental Setup
The model was evaluated on the CL-Drive Dataset, which includes EEG (256 Hz), ECG (512 Hz), and EDA (128 Hz) data from 21 participants in a driving simulator. Tasks were categorized into binary (High vs. Low) and ternary (High/Medium/Low) cognitive load levels. Data augmentation techniques, including Gaussian noise, time warping, and amplitude scaling, were applied to enhance model generalization.
Key Results
UniPhyNet demonstrated superior performance across multiple classification tasks using 10-fold and Leave-One-Subject-Out (LOSO) cross-validation.
- Multimodal Performance: In binary classification (EEG+ECG+EDA), UniPhyNet achieved 80.16% accuracy, significantly outperforming standard raw-signal VGG and ResNet baselines (~70%).
- Unimodal Superiority: On EEG data alone, the model reached 79.29% accuracy, surpassing the widely used EEGNet architecture.
- Generalization: In LOSO testing, the model maintained 73.94% accuracy, indicating that it learns robust physiological markers rather than subject-specific noise.
Ablation studies (shown in Figure 3) highlight that the removal of the GRU layer or data augmentation leads to a noticeable decline in validation accuracy and stability, confirming the importance of temporal modeling and robust training strategies.
Discussion and Conclusion
The success of UniPhyNet underscores the power of end-to-end learning for physiological signals. Key insights include:
- Multiscale Kernels: Combining small and medium kernels prevents overfitting and captures a broader range of features.
- Attention Mechanisms: The CBAM module is critical for filtering noise in physiological data, providing a significant boost in validation performance.
While UniPhyNet is computationally more intensive than simpler models, its ability to handle raw multimodal data with high accuracy makes it a compelling solution for real-world cognitive monitoring. Future work will focus on more granular classification schemes to predict precise cognitive load scores.
Summary updated