TensorPat: Explainable artificial intelligence for EEG-based violence-stimulus classification

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background: EEG responses to violence-related visual stimuli are relevant to neuroscience and digital forensics. Yet most EEG classification models emphasize predictive performance and provide limited evidence about the signal patterns that support each decision. Method: In this research, Tensor Pattern (TensorPat) was proposed as a three-dimensional feature extractor for EEG-based violence-stimulus classification. TensorPat was combined with CWINCA for feature selection, tkNN for classification, and Directed Lobish (DLob) for explainable result generation. In this way, a lightweight and traceable explainable feature engineering framework was developed. Results: A new 32-channel EEG dataset was collected from 34 participants, yielding 527 violence and 1285 control segments. TensorPat generated 10,240 features per segment, of which 131 were retained by CWINCA. LOSO CV was the primary validation protocol. The model achieved 96.14% accuracy and 94.37% balanced accuracy. Violence and control sensitivities were 90.13% and 98.60%, respectively. The DLob sentence had a complexity ratio of 90.92% and identified recurrent frontal, occipital, and parietal channel-symbol patterns. These outputs are model-derived explanations rather than direct activation maps. Conclusions: TensorPat provides a lightweight and explainable framework for EEG-based violence-stimulus classification. DLob outputs should be interpreted as model-derived explanations, not as direct brain activation maps or clinically validated biomarkers. The between-subject acquisition design defines the present scope of inference; matched within-subject cohorts and external datasets are required for broader validation.