HCF-Net: A Hierarchical Co-evolutionary Fusion Network for continuous affective state estimation from synergistic EEG-cardiac dynamics

Objective assessment of affective states is a critical unmet need in managing mood disorders such as major depression, where reliance on subjective self-reports limits timely intervention. To address this, we propose HCF-Net, a multimodal deep learning framework for continuous affective state estimation from the synergistic dynamics of EEG and ECG signals (in DEAP, ECG activity is represented by the plethysmograph-derived blood volume pulse, BVP). The core of our framework is a Hierarchical Co-evolutionary Fusion (HCF) mechanism. Instead of performing only a single interaction at the top layer, HCF inserts layer-wise cross-attention updates at selected encoder depths so that EEG and ECG feature representations can iteratively refine one another during encoding. This interaction is anchored by two specialized frontends: a parameter-free Mutual-Referencing Fusion Core (MRFC) for the enhancement of EEG time–frequency features (DE and PSD), and a lightweight Mamba-based architecture, PhysioMamba, engineered to efficiently capture long-range dynamics in ECG waveforms. To support scientific interpretation, we further introduce an attention- and gate-based analysis protocol to visualize the learned cross-modal interactions. When validated on the public DEAP dataset, HCF-Net achieved a Mean Absolute Error of 1.7609 and a Mean Squared Error of 4.4916. These results suggest that layer-wise multimodal fusion is a promising design choice for continuous affective state estimation from physiological signals. • We propose HCF-Net with cross-attention for EEG-ECG, replacing late fusion. • We introduce PhysioMamba, a Mamba ECG model with CNN frontend for multi-scale, long-range dynamics. • We design MRFC, a parameter-free EEG frontend for DE/PSD; with HCF, it boosts DEAP regression.