Machine Learning Integrating SERF-MEG and VEP for Diagnosis and Differential Diagnosis of Optic Neuropathies

Visual evoked potential (VEP) is widely used to assess optic nerve function, but its ability to differentiate optic neuropathies remains limited. This study investigated the diagnostic value of spin-exchange relaxation-free magnetoencephalography (SERF-MEG), alone and in combination with VEP, using machine learning. A total of 142 eyes, including 71 healthy controls (HC), 34 eyes with optic neuritis (ON), and 37 eyes with ischemic optic neuropathy (ION), were enrolled. Three feature sets were constructed from VEP, SERF-MEG, and their combination. Repeated stratified 5-fold nested cross-validation was used to evaluate nine supervised machine learning algorithms, and SHAP analysis was used to evaluate the interpretability of the models. The multilayer perceptron model obtained an AUC of 0.792 and an MCC of 0.470, whereas the combination VEP-MEG model performed the best in differentiating HC from optic neuropathies. The MEG model fared better than the VEP and combined models for distinguishing ON from ION, with logistic regression obtaining the greatest AUC of 0.847. Cortical network measurements and visually evoked response properties were found to contribute the most to model prediction, according to SHAP analysis. These results imply that SERF-MEG may enhance the diagnosis and differential diagnosis of optic neuropathies by offering supplementary neurophysiological data beyond traditional VEP.