Do peripheral physiological signals decode arousal better than EEG in neurofeedback?

Peripheral signals beat EEG on accuracy, but structure matters more

The study revisited a public dataset from a difficult boundary-avoidance flight task in which participants received EEG-based BCI neurofeedback, sham feedback, or no feedback. The authors trained a multimodal deep learning decoder on seven peripheral channels grouped into four modalities: cardiac (heart rate and HRV-pNN35), respiratory, electrodermal (phasic and tonic), and pupillometric (left and right pupil). Each modality passed through its own two-layer 1D convolutional encoder into a 64-dimensional embedding; the four embeddings were concatenated and classified by a two-layer MLP.

That peripheral decoder reached a within-subject AUC of 93.0% ± 1.1%. A retrained EEG-only variant with the same architecture, applied to 64-channel EEG split into theta, alpha, beta and gamma bands, reached 85.2%. The original filter-bank common spatial pattern (FBCSP) decoder from the earlier study reached 79.8% ± 7.2%. All pairwise differences were significant (p < 0.001, paired t-tests), and all 16 subjects exceeded 80% AUC under the autonomic decoder, with 14 of 16 above 90%.

Adding EEG to the peripheral decoder produced a combined variant at 94.2% ± 0.71%, only 1.2 percentage points above peripheral-only. Integrated Gradients attribution showed cardiac signals dominating predictions, tonic electrodermal activity contributing substantially, respiration and pupil diameter contributing modestly, and EEG attribution essentially zero even in the combined model. The authors read this as evidence that once the autonomic axis is well represented, scalp EEG adds little unique information about graded arousal.

AUC here measures how well the decoder separates trials by the task's arousal surrogate, not whether the decoded number behaves like arousal. That distinction drives the rest of the paper.

Only the peripheral decoder recovers the inverted-U

The Yerkes-Dodson law predicts that performance peaks at an intermediate arousal level and falls off on either side. The authors tested this by fitting linear mixed-effects models of trial performance on decoded arousal, its square, ring difficulty and feedback condition, with subject as a random effect. A significant negative coefficient on the squared term indicates an inverted-U.

For the peripheral decoder, both the linear and quadratic terms were significant (p < 0.01). The estimated optimum sat at roughly 51 on the 0–100 normalized arousal index for both easy and hard courses, with the two curves differing mainly in peak height. For the FBCSP EEG decoder, which the authors re-analyzed themselves rather than taking from the original study, the curve was much shallower and the estimated optima diverged: about 62 on the easy course and about 76 on the hard course.

AIC comparison favored the peripheral quadratic model, with ΔAIC > 18 over the next-best peripheral linear variant and ΔAIC > 44 over both EEG-based models. As an external check, the peripheral decoder correlated with raw physiological markers: r = 0.71 (p < 0.001) with heart rate and r = −0.58 (p < 0.001) with HRV-pNN35, while the EEG decoder correlated substantially more weakly with both.

The authors argue that a decoder can score well by detecting any feature correlated with task difficulty while remaining physiologically uninformative as an arousal index. The drifting FBCSP optimum is consistent with that failure mode, and a control variable with a difficulty-dependent optimum would be miscalibrated in a closed loop.

Feedback did not shift arousal trajectories

Comparing mean arousal traces across BCI, sham and silence conditions, the authors found no significant differences in the time-resolved trajectory (cluster-based permutation, all p > 0.10) or in trial-level summaries such as mean arousal, peak arousal and area under the trajectory (all p > 0.10). Participants produced indistinguishable overall arousal trajectories regardless of which feedback they received.

The authors treat this as substantive rather than a null result. The earlier closed-loop study reported that veridical feedback let subjects fly roughly 7 s longer on the hard course than in silence. If that performance gain occurs without a sustained shift in arousal, the authors reason, the benefit likely comes from localized regulation at performance-critical moments, such as approaching small rings under high demand, rather than from lowering or raising arousal across the whole trial.

This is the authors' interpretation of a pattern in offline data, not a measured mechanism. It also implies that evaluating arousal interventions by mean or peak arousal may miss effects that appear only around specific events.

The optimal arousal level rises across a trial

Using hard-course trials under silence and sham conditions (approximately 240 trials), the authors computed mean performance for each combination of 1-second time bin and 5-unit arousal bin, smoothed the surface, and took the arousal value at peak performance in each time bin as the optimal level. The resulting optimal trajectory was not flat: it rose from low values at trial onset to higher levels by roughly the 45-second mark, tracking the escalating demand as ring sizes decreased.

A fixed threshold applied uniformly across trial time would therefore be miscalibrated for much of each trial. This is the stated basis for making the control band a function of trial time rather than a constant.

Four deviation metrics computed against this trajectory all predicted trial flight time on hard-course control trials: percentage of time above the band (r = −0.28, p = 0.002), percentage of time within the band (r = +0.28, p = 0.002), mean excursion magnitude (r = −0.24, p = 0.010) and excursion rate per minute (r = −0.28, p = 0.002), all Holm-Bonferroni corrected. Top-quartile trials spent about 80% of their duration inside the optimal confidence band; bottom-quartile trials spent about 57% (pairwise contrast p < 0.01).

Baseline HRV and gamma power split subjects into two groups

From a 10-minute calibration session the authors extracted resting HRV-pNN35 and gamma power in the 30–55 Hz band, averaged each across the session, and z-scored across subjects. They formed a sensitivity score s = −HRV_z + Gamma_z, combining low parasympathetic tone with high cortical excitability, and split subjects at the median into arousal-sensitive and arousal-tolerant groups.

Both interaction terms in the mixed-effects model were significant. The HRV_z × arousal² interaction (p = 0.017) indicated that subjects with higher resting HRV had a flatter curve, tolerating a wider arousal range without performance loss. The Gamma_z × arousal² interaction (p = 0.036) indicated that high resting gamma power came with a steeper, narrower peak that drops more sharply at high arousal.

The coupling between deviation and performance also differed by group. Among arousal-sensitive subjects, percentage of time within the optimal band correlated with flight time at r = 0.39 (p = 0.003); among tolerant subjects the relationship was weaker. Restricting to the most extreme cases amplified the gap: r = 0.34 for the sensitive subgroup and r = 0.19 (p = 0.336) for the tolerant subgroup, essentially flat.

The authors read this as an argument against one uniform control band: in the most tolerant individuals, time inside the band carries little predictive value, consistent with a wide acceptable operating range.

Personalized bands improve trial separation, modestly

A grid search over band width multipliers, from 0.50 to 3.00 in steps of 0.25, selected a tight band of 1.5× the standard deviation for the sensitive group and a much wider 2.5× band for the tolerant group. The search jointly maximized Cohen's d between good and bad trials on the percentage-of-time-above-band metric and the Pearson correlation between that metric and performance.

Under a uniform 1.0× multiplier, Cohen's d was 1.21 for the sensitive group and 0.85 for the tolerant group. Under the group-optimized multipliers these rose to 1.32 and 1.10. At the dataset level, the deviation-performance correlation improved from r = 0.308 under the universal band to r = 0.359 under the personalized band.

The authors describe the correlation gain as modest and note it was not subjected to a formal test of the difference between dependent correlations, so it should be read as a directional improvement rather than a statistically established one. The larger gain appeared in the tolerant group, which is most penalized by an overly tight uniform threshold.

The dwell-time analysis bears on whether a reactive intervention could act in time. After excluding single-second threshold crossings, mean dwell time above the personalized band before exit was 6.1 ± 3.6 s for sensitive subjects and 8.6 ± 7.4 s for tolerant subjects, with no sustained excursion below 2 s.

What the offline specification does not show

Every analysis in this paper is offline, and the peripheral decoder was never placed in the feedback loop. The performance advantage of the earlier EEG-based feedback therefore cannot be attributed to improved feedback quality here; establishing that requires a prospective study. The authors sketch one: a preregistered crossover design with active transcutaneous vagus nerve stimulation, sham stimulation and silence, N ≈ 24, triggered when the decoded arousal trajectory exits the personalized band.

The individualization results rest on N = 16 subjects, and the median split leaves n = 8 per sensitivity group. The multipliers 1.5 and 2.5 were selected by grid search within those same small groups and were not validated on held-out data, and the sensitivity classification itself has not been cross-validated as a predictor of the optimal multiplier in an independent sample. The authors present these as proof-of-concept findings and call for replication in at least 30–40 subjects with pre-specified cutoffs.

The decoder was also trained on a single task paradigm using ring size as an arousal surrogate. Whether it generalizes to other arousal sources such as sustained fatigue, emotional stressors or pharmacological interventions is unknown.

The paper's own priority is a comparison, not a threshold: a Yerkes-Dodson-valid arousal signal at 93.0% AUC, an intervention timed to personalized deviations from a time-varying optimal trajectory rather than a fixed threshold, and a stimulation modality matched to moment-targeted regulation. The 1.5× versus 2.5× band split is the concrete decision rule that follows, and it remains a hypothesis awaiting prospective test.

Sources used in this answer