Abstract
Assessing empathy in early childhood is challenging due to the subjectivity of traditional measures like parent surveys and behavioral observation. This paper introduces BEAM (Brainwave Empathy Assessment Model), a deep learning framework that leverages multi-view EEG signals and a pre-trained Transformer (LaBraM) to objectively predict empathy levels in children aged 4 to 6. By integrating cognitive (Theory of Mind) and emotional empathy features through a specialized fusion module and contrastive learning, BEAM significantly outperforms existing state-of-the-art models in identifying a child's prosocial "willingness to help."
Introduction
Empathy is a multi-dimensional construct composed of Cognitive Empathy (Theory of Mind or ToM—understanding another's perspective) and Emotional Empathy (EM—sharing another's feelings). In early childhood, measuring these processes is notoriously difficult; self-reports are unreliable in toddlers, and manual observations are prone to bias. While Electroencephalography (EEG) offers an objective window into the brain, most existing studies focus on adults or ignore the complex temporal dynamics of the developing brain. BEAM addresses these gaps by treating empathy as a multi-view neural process, capturing how ToM and EM interact to produce prosocial behavior.
Methods
The BEAM framework utilizes several sophisticated machine learning techniques to process EEG signals recorded while children watch empathy-evoking stimuli:
- LaBraM Encoder: Instead of training from scratch, the model uses the Large Brain Model (LaBraM), a Transformer-based architecture pre-trained on over 2,500 hours of EEG data. This allows for high-quality feature extraction from relatively small datasets.
- Feature Fusion Module: A key innovation where latent features from ToM and EM views are decomposed into "shared" and "specific" components. A custom fusion loss ($L_{Fusion}$) ensures shared information is integrated while unique characteristics are preserved.
- Contrastive Learning: To mitigate cross-subject noise, the authors applied an InfoNCE loss (Contrastive Alignment). This encourages the model to group similar empathy levels together in the latent space while pushing different classes apart.
Experimental Setup
The study utilized the Chinese Baby Connectome Project (CBCP) dataset, involving 57 typically developing children (average age ~4.9 years). Participants watched the Pixar short film Partly Cloudy to evoke neural responses. Empathy was quantified via a post-test assessment of the child's "willingness to help" in negative scenarios, which served as the ground truth for binary classification (high vs. low empathy).
Key Results and Insights
BEAM demonstrated a significant performance breakthrough compared to existing State-of-the-Art (SOTA) methods. As shown in the comparison below, BEAM achieved an accuracy of 64.7%, markedly higher than BIOT (56.4%) and ST-Transformer (51.2%).
Key Findings:
- ToM as a Primary Driver: Ablation studies revealed that Theory of Mind (ToM) features are stronger predictors of prosocial behavior (61.4% accuracy) than Emotional Empathy (EM) features (58.8%). This suggests that for young children, understanding a situation is a more significant driver of helping behavior than simply feeling distress.
- The Power of Fusion & Contrast: The specialized fusion module and contrastive learning provided measurable boosts in stability and performance, helping the model focus on discriminative features consistent across different children despite individual neural variability.
Limitations and Future Work
While promising, the study is limited by a small sample size (n=57) and the use of a binary "willingness to help" proxy for the complex state of empathy. Future research will focus on developing child-specific encoders, as children's brain signals differ significantly from the adult data used to pre-train models like LaBraM, and expanding the dataset to capture more nuanced social development.
Summary updated