Construction of a diagnostic model for temporal lobe epilepsy using interpretable deep learning: disease-associated markers identification.

INTRODUCTION

Temporal lobe epilepsy (TLE) represents a significant neurological disorder with complex genetic underpinnings. This study aimed to develop an interpretable deep learning diagnostic model for TLE and identify disease-associated markers.

METHODS

Using RNA-seq and microarray data from 287 samples collected from eight GEO datasets, we constructed multiple machine learning algorithms including Deep Neural Networks (DNN), Extreme Gradient Boosting (XGBoost), Random Forest (RF), Logistic Regression (LR), and K-Nearest Neighbors (KNN) to distinguish TLE from normal. SHapley Additive exPlanations (SHAP) and Kolmogorov-Arnold Networks (KAN) were employed to interpret the model and identify key genes associated with TLE pathogenesis.

RESULTS

After comparative analysis, a Deep Neural Network (DNN) model with 10 optimized genetic features achieved perfect diagnostic performance (AUC = 1.000, accuracy = 1.000). SHAP interpretation identified DEPDC5, STXBP1, GABRG2, SLC2A1, and LGI1 as the most significant TLE-associated genes. The KAN model revealed complex nonlinear relationships between these genes and TLE status, providing mathematical expressions that capture their contributions. To facilitate clinical application, we developed an online diagnostic platform that delivers interpretable predictions based on gene expression values.

DISCUSSION: This study advances our understanding of TLE pathogenesis and provides a transparent, interpretable diagnostic model, which combines with traditional diagnostic methods may significantly improve the accuracy of TLE diagnosis, serving as a supplementary tool for clinical assessment.