Abstract
PURPOSE: Clinical trials for remyelination in multiple sclerosis (MS) require an imaging biomarker. The multifocal visual evoked potential (mfVEP) is an accurate technique for measuring axonal conduction; however, it produces large datasets requiring lengthy analysis by human experts to detect measurable responses versus noisy traces. This study aimed to develop a machine-learning approach for the identification of true responses versus noisy traces and the detection of latency peaks in measurable signals. METHODS: We obtained 2240 mfVEP traces from 10 MS patients using the VS-1 mfVEP machine, and they were classified by a skilled expert twice with an interval of 1 week. Of these, 2025 (90%) were classified consistently and used for the study. ResNet-50 and VGG16 models were trained and tested to produce three outputs: no signal, up-sloped signal, or down-sloped signal. Each model ran 1000 iterations with a stochastic gradient descent optimizer with a learning rate of 0.0001. RESULTS: ResNet-50 and VGG16 had false-positive rates of 1.7% and 0.6%, respectively, when the testing dataset was analyzed (n = 612). The false-negative rates were 8.2% and 6.5%, respectively, against the same dataset. The latency measurements in the validation and testing cohorts in the study were similar. CONCLUSIONS: Our models efficiently analyze mfVEPs with <2% false positives compared with human false positives of <8%. TRANSLATIONAL RELEVANCE: mfVEP, a safe neurophysiological technique, analyzed using artificial intelligence, can serve as an efficient biomarker in MS clinical trials and signal latency measurement.