Artificial neural network based automatic detection of motor evoked potentials

Introduction: Motor evoked potentials (MEP) are detected using various methods that determine signal changepoints. The current detection methods perform well given a high signal to noise ratio. However, performance can diminish with artefact such as those arising due to poor signal quality and unwanted electrical potentials. Part of the problem is likely because the methods ignore the morphology of a signal making it impossible to differentiate noise from MEPs. Methods: For the first time, we investigated a new detection method able to learn MEP morphology using artificial neural networks. To build an MEP detection model, we trained deep neural networks with architectures based on combined CNN and LSTM or self-attention mechanism, using sample MEP data recorded from able-bodied individuals. The MEP detection capability of the models was compared with that of a changepoint based detection method. Results: Our models reached test accuracy of up to 89.7 ± 1.5 % on average. In a real-world setting evaluation, our models achieved average detection accuracy of up to 94.7 ± 1.2 %, compared with 76.4 ± 5.3 % for the standard changepoint detection method (p = 0.004). Conclusion: Artificial neural network models can be used for improved automated detection of MEPs.