ST-AFNet: A spatio-temporal-attentional fusion network for mindfulness meditation state recognition using biosignals

Mindfulness meditation state recognition faces critical challenges, including multimodal signal heterogeneity, ambiguous state boundaries, and limited individual generalization. To address these, we propose the Spatio-Temporal-Attentional Fusion Network (ST-AFNet), a dynamic framework integrating electroencephalogram (EEG), electrocardiogram (ECG), and galvanic skin response (GSR) biosignals. Our approach establishes a 500ms-resolution dynamic brain network using phase-locking matrices and introduces a gated fusion mechanism to adaptively weight multimodal contributions—revealing EEG dominance (82%) during focused states and elevated ECG input (38%) during relaxation. Coupled Bidirectional LSTM-CRF (BiLSTM-CRF) modeling enforces physiologically plausible state transitions. Validation demonstrates >85% accuracy at signal-to-noise ratios (SNR) > -5 dB, with significant correlations to alpha-band power (r = 0.87, p < 0.001) and high-frequency heart rate variability (HF-HRV; r = 0.83, p < 0.001). This work delivers an interpretable, robust deep learning framework for advancing digital mental health therapeutics.