Metaplastic-EEG: Continuous Training on Brain-Signals

Abstract

AbstractDeep learning approaches promise viable solutions for detecting epileptic seizures in a reliable, generalisable and potentially real-time. To apply such techniques in clinical settings, where they can be used with long-term recordings or applied to a continuous stream of incoming datasets, these algorithms should adopt a continual learning ability that allows the agent to acquire and adapt from additional knowledge streamed over its lifespan. Unfortunately, traditional sequential learning can initiate catastrophic forgetting, in which the model loses previously learned information while accumulating new knowledge. Metaplasticity has emerged as a potential technique to provide longer-term stability pertaining to the learning performance for multiple datastream sets, thus enabling a meta-learning capability in artificial learning machines and algorithms. In this paper, we use these biologic-inspired metaplasticity techniques to develop stable learning cycles when we expose it to multiple sets of EEG (electroencephalogram) data for seizure detection. In this feasibility study, adding metaplastic synapses enhances detection accuracy relative to traditional baseline learning. Considering the meta-learning approach demonstrated in this paper, metaplastic binarized neural networks (BNNs) demonstrate improvement (6-7%) in seizure detection performance metrics, with reported accuracies and ROC-AUC values over 70%. Metaplasticity in practice with machine learning holds the potential to provide an adaptable, patient-specific epileptic seizure tracking method for real-world dynamics.