Figure 1: Multimodal peripheral decoder architecture and performance. (A) Network architecture: each input modality (EEG, EDA, HR/HRV, RESP, Pupil) is independently embedded by a Conv1D+BatchNorm temporal encoder, embeddings are concatenated, and a two-layer MLP classifier produces the arousal index. (B) Within-subject AUC for the original FBCSP benchmark, a retrained EEG-only decoder, the autonomic-only decoder, and the combined autonomic+EEG decoder. Error bars are SEM across subjects. (C) Integrated Gradients feature attribution by modality across the 2-second pre-event epoch. Cardiac signals (HR) dominate; EDA contributes substantially; respiration shows a late-epoch rise.## 4 Results 出處: Multimodal Physiological Decoding Reveals Individualized Arousal Dynamics in Closed-Loop Neurofeedback（arXiv）Figure 4: Optimal arousal trajectory and performance prediction. (A) Heatmap of mean flight time as a joint function of decoded arousal (y-axis) and trial time (x-axis), computed from hard-course control trials; warmer colors indicate lower mean flight time. The blue dashed line is the polynomial-fit optimal trajectory derived from the row of peak performance at each time bin. (B) Trial-level scatter of flight time versus % time within the optimal band (r=0.308r=0.308, p<0.001p<0.001). (C) Mean arousal traces (±\\pm SEM) for top-quartile (green), middle (gray), and bottom-quartile (red) trials, overlaid on the optimal trajectory (black dashed) and ±1\\pm 1 SD confidence band (gray shading). Bad trials rise above the optimal trajectory early; good trials track it closely from below. (D) % time within optimal band by performance tertile. Good trials spend ∼80%\\sim 80\\% in band, bad trials ∼57%\\sim 57\\% (p<0.01p<0.01, Holm-corrected).### 4.6 Individual differences in deviation-performance coupling 出處: Multimodal Physiological Decoding Reveals Individualized Arousal Dynamics in Closed-Loop Neurofeedback（arXiv）
