Neuroscore (EEG): Neural Response Metrics

• Neuroscore (EEG) is a family of empirical quantifiers that measure neural processing and perceptual quality using EEG signals.
• One approach computes the mean single-trial P300 amplitude through spatial beamforming, latency optimization, and covariance pooling to index psychoperceptual salience.
• An alternative method employs q-statistical mechanics applied to EEG inter-event intervals to quantify global and local neural complexity, correlating with cognitive state and aging.

Neuroscore (EEG) is a family of empirical quantifiers for neurophysiological responses, derived from electroencephalogram (EEG) time series, and designed to capture information about neural processing, perceptual quality, or brain complexity. Two prominent forms of Neuroscore have emerged in the literature: one based on P300 amplitude as an index of psychoperceptual salience, and one based on $q$-statistical complexity. Both approaches utilize EEG signals but differ fundamentally in their computational definitions, neuroscientific interpretations, and methodological requirements.

1. Neuroscore via P300 Amplitude: Definition and Computational Protocol

Neuroscore, as originally formulated for image evaluation, is defined as the mean single-trial P300 amplitude evoked by a specific image class (e.g., GAN-generated faces, real faces) under a rapid serial visual presentation (RSVP) paradigm. Its computation involves:

• EEG epoch extraction: Let $X_i\in\mathbb{R}^{C\times T}$ denote the multichannel EEG epoch for the $i$th “target” trial (images from a given class), and $K_j\in\mathbb{R}^{C\times T}$ for the $j$th standard (non-target) trial, with $C$ as channel count and $T$ temporal resolution.
• Covariance pooling: Compute a pooled spatial covariance matrix:

$$\Sigma = \frac{1}{N}\sum_{i=1}^N X_iX_i^\top + \frac{1}{M}\sum_{j=1}^M K_jK_j^\top \in \mathbb{R}^{C\times C}$$

• Spatial beamforming: For each candidate latency $t$ (within the post-stimulus window, typically 400–600 ms), define the spatial contrast

$$p(t) = \frac{1}{N}\sum_i X_i(:,t) - \frac{1}{M}\sum_j K_j(:,t).$$

Solve the constrained minimization

$$\min_w\, w^\top\Sigma w,\quad \text{s.t.}\; w^\top p(t) = 1.$$

yielding closed-form solution

$$w(t) = \frac{\Sigma^{-1} p(t)}{p(t)^\top \Sigma^{-1} p(t)}.$$

• Optimal latency selection: Choose $t^*$ that minimizes $w(t)^\top\Sigma w(t)$. Use $w^* = w(t^*)$ for dimensionality reduction.
• Single-trial source waveforms: For each trial, project as $s_i = {w^*}^\top X_i \in \mathbb{R}^T$.
• Peak quantification: Extract peak amplitude $$A_i = \max_{\tau\in[t^*-100\,\text{ms},\,t^*+100\,\text{ms}]} s_i(\tau)$$.
• Aggregate metric: Compute

$\text{Neuroscore} = \frac{1}{N}\sum_{i=1}^N A_i$

as the index for the considered image class (Wang et al., 2019).

This method explicitly leverages task-evoked ERPs, particularly the P300, as a marker of conscious visual detection and psychoperceptual salience.

2. $q$-Statistical Neuroscore: Theoretical Basis and Mathematical Construction

The $q$-statistical Neuroscore applies nonextensive statistical mechanics (NESM), generalizing Boltzmann–Gibbs–Shannon statistics to systems exhibiting nonadditive entropy and long-range correlations. Here, complexity is indexed by the entropic parameter $q$, estimated from EEG inter-event interval distributions.

• Empirical model: Let $x$ be the inter-event interval (e.g., successive negative crossings below $-1\,\mathrm{SD}$). The distribution

$$P(x) = \frac{a x^c}{[1 + (q-1)b x^h]^{1/(q-1)}},\quad q>1,\,b>0,\,h>0$$

with normalization constant $a$ as

$$a^{-1} = [b(q-1)]^{-\frac{c+1}{h}} \frac{\Gamma\left(\frac{c+1}{h}\right) \Gamma\left(\frac{1}{q-1} - \frac{c+1}{h}\right)}{h \Gamma\left(\frac{1}{q-1}\right)}.$$

recovers the standard exponential for $q\to 1$. The “fatness” of the tail, controlled by $q$, is interpreted as a marker of neural complexity (Abramov et al., 9 Feb 2025).

• Parameter estimation: Nonlinear least squares is used to fit the parameters $(b, c, h, q)$ to EEG interval histograms, with $q$ serving as the complexity Neuroscore.

This approach yields global (whole-sensor) and local (per-sensor) $q$ indices quantifying hierarchical, non-local system complexity and its modulation by cognitive state, neurodevelopment, or pathology.

3. Experimental Protocols and EEG Preprocessing

P300-based Neuroscore (Wang et al., 2019):

• Participants: 12 adults in alternating blocks of behavioral discrimination and RSVP EEG.
• Stimulus presentation: RSVP block with 240 images at 4 Hz (250 ms/image); target images drawn from DCGAN, BEGAN, PROGAN, and real faces; standards are non-faces.
• EEG recording: 32-channel cap, 1 kHz sampling, photodiode-aligned.
• Preprocessing pipeline:
  • CAR re-reference.
  • 0.5–20 Hz band-pass filter.
  • Downsample to 250 Hz.
  • 0–1 s epoch extraction.
  • Artifact rejection by amplitude threshold.
  • Retain only trials with valid behavioral response (0–1 s).

$q$-statistical Neuroscore (Abramov et al., 9 Feb 2025):

• Participants: 70 adults in seven defined functional states (eyes open, eyes closed, math, music, etc.).
• EEG recording: 20 scalp electrodes, 1 kHz sampling.
• Preprocessing: 0.5–100 Hz band-pass, 60 Hz notch, artifact removal (>3 SD and ECG/manual).
• Event definition: Downward crossings below $-1$ SD; epochs truncated at ±100 μV.
• Interval histogram: 0–1000 ms range, 2 ms bins, exclude 8–12 Hz rhythms.
• Curve fitting: Simultaneous estimation of $(b, c, h, q)$, separate for each channel and globally.

Both protocols entail rigorous artifact control, frequency-domain filtering, and optimal alignment between stimulus and EEG stream.

4. Empirical Evaluation and Quantitative Benchmarks

Human-Consistency and Perceptual Alignment (P300-based):

• Strong negative correlation between real Neuroscore and behavioral discriminability: $r(36) = -0.828$, $p=4.77\times10^{-10}$—larger P300 amplitude marks greater perceived realism in images.
• Synthetic-Neuroscore prediction error (across three GANs, per subject; mean ± SD):
  • Shallow-EEG: $0.151 \pm 0.245$
  • Shallow(no EEG): $0.428 \pm 0.623$
  • MobileNet-EEG: $0.155 \pm 0.235$
  • MobileNet(no EEG): $0.437 \pm 0.589$
  • Inception-EEG: $0.157 \pm 0.487$
  • Inception(no EEG): $0.462 \pm 0.932$
• Inclusion of real EEG during training reduces prediction error by a factor of 2–3 compared to no-EEG or randomized-EEG controls.
• GAN ranking by synthetic-Neuroscore (with EEG) matches human ranking and outperforms traditional Inception Score, MMD, and FID for alignment.

Complexity, Functional State, and Individual Factors ($q$-statistical):

• Global vs. local complexity: $q_{\mathrm{AllCh}}$ (global) consistently exceeds mean $q$ over channels (local), e.g. Rest OE: $1.196\pm0.053$ vs $1.176\pm0.047$ ($t=2.75$, $p=0.008$).
• Correlation with EEG bands: $q_{\mathrm{AllCh}}$ correlates positively with $\theta$ and $\alpha$, negatively with $\beta_1$ (e.g., $\theta$: $r=+0.48$, $p=1\times 10^{-4}$; $\beta_1$: $r=-0.25$, $p=0.057$).
• Functional state modulation: Local $q$ exhibits site- and task-specific modulation (e.g., reduced occipito-parietal $q$ during eyes closed, increased right-lateral posterior $q$ during music).
• Age effects: Negative correlation between age and $q_{\mathrm{AllCh}}$ (resting OE: $r=-0.50$, $p<10^{-4}$).

5. Neuro-AI Integration and Model Architectures

In the P300 framework, a convolutional neural network (CNN) “neuro-AI interface” is trained to predict Neuroscore from the image alone:

• Model backbone: Shallow Net (custom), MobileNet V2, or Inception V3 (pretrained).
• Fully connected layers: Multi-stage, culminating in waveform prediction (size $T$) and scalar amplitude output.
• Loss definition:
  • Stage 1: Minimize

  $$\mathrm{loss}_1(\theta_1) = \frac{1}{N}\sum_{i=1}^N \| S_i^{\mathrm{true}} - S_i^{\mathrm{pred}}(\theta_1) \|_2^2$$

  for waveform reconstruction. - Stage 2: Freeze $\theta_1$, minimize

  $$\mathrm{loss}_2(\theta_1,\theta_2) = \frac{1}{N}\sum_{i=1}^N (y_i^{\mathrm{true}} - y_i^{\mathrm{pred}}(\theta_1,\theta_2))^2$$

  for amplitude regression.

• Training regime: 20 epochs, batch size 256, Adam optimizer (lr = 0.001), fine-tuning on FC layers only.
• Ablations: Exclusion or randomization of EEG signals degrades predictive alignment with human judgment.

This pipeline enables the quantitative translation of EEG-based P300 response into a synthetic-Neuroscore for image (e.g., GAN output) quality assessment.

6. Practical Guidelines and Limitations

Feature	P300-based Neuroscore	$q$-statistical Neuroscore
Sample size (per class)	$\geq$20–30 trials (P300 saturates)	Large continuous recordings
Channels	32 (RSVP paradigm)	20 (10–20 system)
Core metric	Mean P300 amplitude	$q$ from interval distribution
Predictive interface	CNN (with/without EEG supervision)	Parameter estimation only
Sensitivity	Perceptual/psychometric realism	Intrinsic/global complexity

• P300-based: Multiple clean trials per class necessary; spatial filter limited by $C\lesssim64$; computationally efficient (real-time feasible); generalizes if sufficient EEG-labeled images are available; requires RSVP paradigm for P300 elicitation (Wang et al., 2019).
• $q$-statistical: Event definition and artifact rejection critical; suppression of dominant rhythms required; spatial detail constrained by 20-channel density; fit is sensitive to outliers and poorly convergent cases (~5% excluded) (Abramov et al., 9 Feb 2025).

7. Significance and Applications

Neuroscore offers reproducible, interpretable EEG-derived metrics with distinct neuroscientific significance:

• P300-derived: Aligns with human perceptual evaluation of generated visual content; enables deep learning models to internalize explicit neural indices of quality, outperforming traditional image metrics in human consistency.
• $q$-complexity-derived: Quantifies the nonadditive complexity architecture of neural dynamics; sensitive to both global integrative states and local/reversible task effects; suitable as a biomarker in cognitive aging, neurodevelopment, and neuropsychiatric evaluation.

Both frameworks represent physiologically grounded, low-variance, noninvasive approaches for translating neural signals into objective “Neuroscores,” providing a bridge between machine learning, psychophysics, and systems neuroscience (Wang et al., 2019, Abramov et al., 9 Feb 2025).