Abstract
Psychological resilience—the capacity to maintain mental health in the face of adversity—is a critical trait that is traditionally difficult to measure objectively. While neuroimaging offers a potential solution, the field is hampered by data scarcity and low signal-to-noise ratios. This paper introduces a novel framework utilizing Neuro-Kolmogorov-Arnold Networks (Neuro-KAN) to predict resilience by integrating EEG and fMRI data. By employing a "smart chunking" strategy and trait-informed contrastive learning, the proposed model outperforms existing state-of-the-art methods, achieving high accuracy on both private and public datasets while providing interpretable insights into the neural hubs of resilience.
Introduction
Predicting psychological resilience is vital for high-stakes professions, such as aviation or emergency services, where self-reported questionnaires are often subject to bias or intentional manipulation. Neurological markers from electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) provide objective alternatives, but they present significant computational challenges. Neuroimaging datasets are typically small, the data is often contaminated by physiological noise, and resilience itself is a distributed trait rather than a localized brain function.
To address these hurdles, this research proposes a data-efficient framework that leverages the unique properties of Kolmogorov-Arnold Networks (KANs) and a multimodal alignment strategy to extract stable resilience traits from noisy, limited neurological signals.
Methodology
Neuro-KAN Architecture
Unlike standard Multi-Layer Perceptrons (MLPs) that use fixed activation functions on nodes, Kolmogorov-Arnold Networks (KANs) utilize learnable functions on the edges (splines). This makes them highly parameter-efficient and particularly suited for scientific data where interpretability is paramount. The framework employs modality-specific backbones:
- EEG Branch: Uses time-wise and electrode-wise convolutions to capture high-resolution temporal events.
- fMRI Branch: Utilizes Graph Convolutional Networks (GCN) to model functional connectivity across 268 brain regions.
Data Augmentation and Alignment
To combat data scarcity, the authors implemented a "smart chunking" strategy, splitting long recordings into smaller segments. The optimal chunk size was determined using Tree-structured Parzen Estimators (TPE) to balance sample quantity with trait stability.
Inspired by contrastive learning models like CLIP, the framework aligns EEG, fMRI, and behavioral data in a shared latent space. A modified loss function ensures that different time chunks from the same individual are treated as positive pairs, reinforcing the model's focus on stable traits rather than transient states. Finally, a Noise-Informed Inference module weights predictions based on the similarity between EEG and fMRI representations; if the modalities disagree, the chunk is flagged as noisy and its influence on the final score is reduced.
Key Results and Interpretability
The model was evaluated on two datasets: RESIL (a self-constructed dataset of 45 subjects) and LEMON (a public dataset of 227 subjects). The proposed architecture significantly outperformed baselines like EEGNet and BrainGNN, achieving an accuracy of 0.8444 on RESIL and 0.7897 on LEMON. In regression tasks, it reached an R² of 0.4185, surpassing the next best multimodal model (CLIP) by a wide margin.
Beyond predictive performance, the use of KANs and GCNs allowed for high interpretability. The researchers identified specific neural correlates of resilience, finding that the right lateral parietal cortex and connections between the inferior parietal lobule and lateral temporal cortex serve as key hubs. These findings align with psychological theories regarding the brain's executive control and emotional regulation networks.
Conclusion
This work demonstrates that psychological resilience can be accurately predicted from neurological data, even when samples are limited. By combining the parameter efficiency of KANs with noise-aware multimodal learning, the framework provides a robust tool for objective mental health assessment. Future research will focus on using the symbolic nature of KANs to derive explicit mathematical formulas for resilience, potentially bridging the gap between deep learning and neurological theory.
Summary updated