Automated detection of trigeminal neuralgia using multi-domain EEG feature analysis and CNN-attention architecture: a comparative machine learning study.

Trigeminal neuralgia (TN) is a debilitating neuropathic pain disorder characterized by sudden, intense facial pain, with diagnosis heavily reliant on subjective symptom reporting, leading to frequent misdiagnosis and treatment delays. This study aimed to develop and evaluate a proof-of-concept electroencephalography (EEG)-based automated detection framework for TN through systematic comparison of multiple classification architectures using multi-domain feature analysis. Resting-state EEG recordings were acquired from 72 subjects (36 TN patients: mean age 44 ± 13.98 years; 36 healthy controls: mean age 27.75 ± 3.15 years) using 19 scalp channels sampled at 250 Hz during 5 min eyes-closed sessions. Following standardized preprocessing, signals were decomposed into five frequency bands using discrete wavelet transform, and 40 quantitative features spanning temporal, spectral, complexity, statistical, and nonlinear domains were extracted per band, generating 3,800 initial features. A two-stage selection procedure identified 28 optimal features with high discriminative power (p < 0.05; |d|> 0.5 for 22 features) and low inter-feature correlation. Six architectures were evaluated through subject-wise fivefold stratified cross-validation: Support Vector Machine, Random Forest, Decision Tree, Deep Neural Network, Baseline Convolutional Neural Network, and a novel Convolutional Neural Network with Attention model employing 64 × 64 grayscale image transformation of multi-domain temporal features. Statistical analysis identified 1,877 significant feature-channel-band combinations, with the Gamma band (30-50 Hz) demonstrating the highest discriminative power (512 significant features; Cohen's d > 5.9 for frontal Approximate Entropy) and frontal regions showing predominant alterations (43.3% of significant features). Analysis of Covariance confirmed that 22 of 28 discriminative features (78.6%) retained significance independent of the inter-cohort age difference (η2p ≥ 0.06). Under subject-wise cross-validation, the CNN with Attention model achieved the highest preliminary performance, with 96.44 ± 2.60% accuracy, ROC-AUC of 0.996 ± 0.007, and recall of 99.8 ± 0.3%, outperforming Random Forest (92.30%), Baseline CNN (93.81%), Deep Neural Network (88.74%), Decision Tree (83.03%), and Support Vector Machine (62.96%). These metrics should be interpreted with caution, as the healthy control cohort comprised only 14 real independent subjects prior to augmentation; consequently, each test fold in the fivefold cross-validation was evaluated against 2-3 truly independent controls alongside their synthetic counterparts, which substantially constrains the generalizability of the reported figures. This study provides preliminary proof-of-concept evidence that EEG-based automated detection is a promising non-invasive approach for TN, with the CNN with Attention architecture demonstrating encouraging performance on a small, augmentation-dependent dataset. The identified neurophysiological signatures, particularly Gamma-band entropy alterations and frontal complexity measures, advance understanding of central nervous system reorganization in TN. However, the reported performance metrics are contingent on a limited number of truly independent controls and must not be interpreted as indicative of clinical readiness. External validation on large, independently acquired, age-matched cohorts is an essential prerequisite before any conclusions regarding diagnostic generalizability can be established.