Interpretable Fuzzy Modeling Reveals Population-Level Representation Differences in P300 Brain Computer Interfaces Across Neurodivergent and Neurotypical Cohorts

1. Problem Statement

Brain-Computer Interfaces (BCIs) based on the P300 event-related potential are widely used for communication, particularly for individuals with motor impairments. However, a major barrier to practical deployment is population heterogeneity.

• The Gap: Existing research primarily focuses on decoding performance (accuracy) or signal-level differences (e.g., ERP amplitude/latency). There is a lack of understanding regarding the representation-level structure learned by decoders across different populations.
• The Challenge: Neurodivergent populations (specifically Autism Spectrum Disorder [AUT] and Amyotrophic Lateral Sclerosis [ALS]) exhibit different sensory processing, attention, and cognitive burdens compared to Neurotypical (NT) cohorts. This leads to variability in P300 morphology. Current "black-box" deep learning models often fail to generalize across these groups and cannot explain why performance degrades or differs, obscuring whether the issue is merely noise or a fundamental reorganization of the discriminative features.

2. Methodology

The authors propose iFuzzyP300, an interpretable fuzzy spatiotemporal framework designed to decode P300 signals while explicitly modeling population-specific representation structures.

A. Data Sources

The study utilized a unified framework to analyze three distinct cohorts:

1. ALS Cohort: Subsets from the bigP3BCI dataset (Studies L and N), involving 19 participants across 1–2 sessions.
2. Autism (AUT) Cohort: The BCIAUT-P300 benchmark, involving 15 participants across 7 sessions (105 total sessions).
3. Neurotypical (NT) Cohort: Subsets from bigP3BCI (Studies E, Q, S1, S2), involving 78 participants.

B. Model Architecture

The model consists of three main components:

1. Spatial Fuzzy Module: Processes EEG channels to learn cohort-specific spatial contribution patterns. It uses learnable fuzzy prototypes to weight channel inputs.
2. Temporal Fuzzy Module: Processes the time dimension to capture event-related waveform structures. It uses Gaussian prototype functions to compute rule activations based on similarity to learned temporal centers.
3. Classification Head: A lightweight Multi-Layer Perceptron (MLP) that outputs class probabilities based on the aggregated fuzzy rule activations.

C. Interpretability Mechanism

The core innovation is Prototype Reconstruction:

• The model learns fuzzy prototypes (centers) in a latent space.
• These latent prototypes are projected back to the original signal domain (using Moore-Penrose pseudoinverse if necessary) to reconstruct waveform patterns.
• This allows researchers to visualize exactly what "fuzzy rule" (e.g., a specific P300 shape) the model has learned for each cohort, rather than just seeing a final accuracy score.

D. Evaluation Protocol

• Within-Subject Setting: Models were trained and tested within the same subject (80% train/20% validation split) to evaluate personalized decoding capabilities.
• Metrics: Accuracy, F1-score, Precision, and Recall were used to address class imbalance (target vs. non-target trials).
• Baselines: Compared against state-of-the-art deep learning models including EEGNet, DeepConvNet, ShallowConvNet, Conformer, and EEGformer.

3. Key Contributions

1. Unified Comparative Framework: Provided the first direct comparison of P300 decoding across ALS, AUT, and NT cohorts using a single, consistent fuzzy modeling framework.
2. Beyond Performance Metrics: Moved beyond aggregate accuracy to analyze learned fuzzy centers as interpretable population-level representations.
3. Representation-Level Insight: Demonstrated that reduced performance in neurodivergent cohorts is not just due to lower signal-to-noise ratio, but is accompanied by systematic differences in the geometry and morphology of the learned discriminative prototypes.

4. Key Results

A. Decoding Performance

• The proposed fuzzy model achieved competitive performance across all cohorts, often outperforming or matching deep learning baselines.
• Stability in Neurodivergent Groups: The fuzzy model showed particular robustness in the AUT cohort, where models like EEGNet and EEGformer exhibited significant performance degradation or instability (e.g., EEGformer dropped to ~33% accuracy).
• ALS Performance: The fuzzy model performed comparably to strong convolutional baselines in ALS studies.

B. Representation Analysis (The Core Finding)

Analysis of the reconstructed fuzzy centers revealed systematic differences:

• Waveform Morphology: Reconstructed centers for the AUT cohort showed visibly larger deviations in waveform shape compared to ALS and NT, particularly in the post-stimulus interval. ALS and NT centers were more similar but still distinct.
• Statistical Significance: Point-wise statistical analysis (ANOVA) confirmed significant temporal differences between cohorts, including intervals overlapping with the canonical P300 window.
• Geometric Separation: Low-dimensional embeddings of the reconstructed centers showed that the fuzzy prototypes for AUT, ALS, and NT formed partially separated clusters. This indicates that the model learns distinct "decision boundaries" and prototype organizations for each population.

5. Significance and Implications

• Interpretability in BCI: The study validates fuzzy logic as a powerful tool for BCI, offering transparency that black-box deep learning models lack. It proves that models can be both accurate and explainable.
• Redefining Population Heterogeneity: The findings suggest that population differences in P300-BCI are not merely "noise" to be filtered out but represent fundamental reorganizations of neural representation. A model trained on NT data may fail on AUT/ALS data not just because of noise, but because the underlying "fuzzy rules" for what constitutes a "target" signal are geometrically different.
• Clinical and Translational Impact:
  • Design: Future P300-BCI systems must be "population-aware," potentially requiring cohort-specific calibration or transfer learning strategies that account for these representation shifts.
  • Neurodiversity: The results support the view that neurodivergent individuals utilize qualitatively different cognitive styles rather than simply exhibiting deficits. This has implications for designing inclusive assistive technologies and workplace practices that accommodate diverse cognitive processing styles.

In conclusion, this paper establishes that interpretable fuzzy modeling is essential for understanding the true nature of P300 variability across neurodivergent and neurotypical populations, bridging the gap between neural signal analysis and robust, clinically meaningful BCI design.