prm-explainable-rnn-p300-bci - SKILL.md Agent Skill

name: prm-explainable-rnn-p300-bci description: "Post-Recurrent Module (PRM) for explainable RNN-based P300 classification in BCIs — combines performance improvement with global/local explainability techniques for transparent EEG-based neural decoding. Activation triggers: PRM, P300 BCI, explainable RNN, EEG explainability, post-recurrent module, P300 classification, transparent BCI."

PRM: Explainable RNN for P300-based Brain-Computer Interfaces

Post-Recurrent Module (PRM) that enhances both performance and transparency of RNN architectures for P300 signal classification from EEG, enabling dual global/local explainability analysis aligned with neurophysiological P300 descriptions.

Metadata

Source: arXiv:2605.10121
Authors: Christian Oliva, Vinicio Changoluisa, Francisco B Rodríguez, Luis F Lago-Fernández
Published: 2026-05-11

Core Methodology

Problem Statement

P300-based Brain-Computer Interfaces face two key challenges:

Inter- and intra-subject variability limits practical deployment
DL model explainability gap prevents clinical/assistive adoption — users need to understand why a model makes certain decisions

Current approaches treat classification as a black box, losing the connection between model decisions and established neurophysiology of the P300 ERP.

Key Innovation: Post-Recurrent Module (PRM)

The PRM is an additional layer incorporated into RNN architectures that:

Improves classification performance (9% over SOTA)
Enables dual explainability analysis:
- Global explainability: Identifies most relevant brain regions and critical time intervals
- Local explainability: Interprets individual decisions in terms of spatio-temporal EEG patterns consistent with P300 neurophysiology

Architecture

Multi-channel EEG input
    │
    ▼
┌─────────────────┐
│  RNN Backbone   │  ← Captures temporal dynamics of EEG signals
│  (LSTM/GRU)     │     across the P300 time window (~250-500ms)
└────────┬────────┘
         ▼
┌─────────────────┐
│  Post-Recurrent  │  ← Additional processing layer
│  Module (PRM)   │     • Enhances spatio-temporal feature representation
│                 │     • Enables attention/weight visualization
│                 │     • Maps features back to spatial (channel) and
│                 │       temporal (time point) dimensions
└────────┬────────┘
         ▼
┌─────────────────┐
│  Classification  │  ← P300 vs non-P300 decision
│  Head           │
└─────────────────┘

Explainability Techniques

Global Explainability
- Aggregate PRM weights/importance across all samples
- Generate spatio-temporal importance maps:
  - Spatial: Which EEG channels (brain regions) contribute most?
  - Temporal: Which time intervals within the epoch are critical?
- Validate against known P300 neurophysiology (centro-parietal positivity at ~300ms)
Local Explainability
- Per-sample importance attribution
- Identify which spatio-temporal patterns drove each specific classification decision
- Enable subject-specific analysis of P300 variability

Experimental Results

9% performance improvement over state-of-the-art methods
PRM importance maps align with established P300 neurophysiology:
- Centro-parietal electrodes show highest importance
- Temporal importance peaks in the 250-500ms post-stimulus window
Captures both inter-subject and intra-subject variability patterns
Results consistent with neuroscience literature on P300 generators

Implementation Guide

Prerequisites

PyTorch
EEG data with P300 paradigm (oddball task)
Standard preprocessing: filtering, epoching, baseline correction

Step-by-Step

Base RNN Construction

import torch
import torch.nn as nn

class PRM(nn.Module):
    def __init__(self, rnn_hidden, n_channels, n_timepoints):
        super().__init__()
        # Spatial attention over channels
        self.spatial_attn = nn.Linear(n_channels, n_channels)
        # Temporal attention over time points
        self.temporal_attn = nn.Linear(n_timepoints, n_timepoints)
        # Fusion layer
        self.fusion = nn.Linear(rnn_hidden * 2, rnn_hidden)

    def forward(self, rnn_output, eeg_input):
        # rnn_output: (batch, hidden)
        # eeg_input: (batch, channels, time)

        # Compute spatial importance
        spatial_weights = torch.softmax(self.spatial_attn(eeg_input.mean(-1)), dim=-1)
        # Compute temporal importance
        temporal_weights = torch.softmax(self.temporal_attn(eeg_input.mean(1)), dim=-1)

        # Weighted feature enhancement
        spatial_enhanced = (eeg_input * spatial_weights.unsqueeze(-1)).mean(1)
        temporal_enhanced = (eeg_input * temporal_weights.unsqueeze(1)).mean(-1)

        # Concatenate with RNN output and fuse
        enhanced = torch.cat([rnn_output, spatial_enhanced, temporal_enhanced], dim=-1)
        return self.fusion(enhanced), spatial_weights, temporal_weights

Training Loop with PRM

model = P300RNNWithPRM(n_channels=32, n_timepoints=250, rnn_hidden=128)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)

for epoch in range(n_epochs):
    for eeg, labels in dataloader:
        preds, spatial_w, temporal_w = model(eeg)
        loss = nn.BCEWithLogitsLoss()(preds, labels)
        loss.backward()
        optimizer.step()

Extracting Explainability Maps

# Global explainability: average attention weights across dataset
all_spatial = []
all_temporal = []
for eeg, _ in dataloader:
    _, sw, tw = model(eeg)
    all_spatial.append(sw.mean(0).detach())
    all_temporal.append(tw.mean(0).detach())

global_spatial_map = torch.stack(all_spatial).mean(0)  # (n_channels,)
global_temporal_map = torch.stack(all_temporal).mean(0)  # (n_timepoints,)

# Plot: spatial map on EEG topography
# Plot: temporal map aligned to P300 time window

Local Explainability for Individual Samples

def explain_single_trial(model, eeg_sample):
    _, spatial_w, temporal_w = model(eeg_sample.unsqueeze(0))
    return {
        'channel_importance': spatial_w[0].detach().cpu().numpy(),
        'time_importance': temporal_w[0].detach().cpu().numpy(),
    }

Applications

P300 speller BCIs: Transparent classification for assistive communication
Motor imagery: Generalize PRM architecture to MI classification
SSVEP: Apply to steady-state visual evoked potential detection
Cognitive workload assessment: Use explainability to identify neural markers of workload
Clinical validation: Verify model decisions align with clinical EEG interpretations

Pitfalls

Subject variability: PRM importance patterns vary significantly across subjects; individual calibration may be needed
EEG montage dependency: Spatial importance maps depend on electrode placement; results may not transfer between montages
Temporal resolution: PRM temporal importance is bounded by EEG sampling rate and epoch length
Over-interpretation: Attention weights indicate feature importance but don't establish causal relationships
Generalization: While PRM generalizes across EEG tasks, the specific importance patterns are task-dependent

Related Skills

eeg-foundation-model-adapters
explainable-gnn-eeg-neurological
bandroutenet-eeg-artifact-removal
eeg-brain-connectivity-bci
copilot-assisted-second-thought-bci
eeg-ieeg-bridge-bci