Towards robust seizure type classification via curriculum learning paradigms

Seizure often manifests in various forms across diverse patient groups. Various machine learning (ML) models have been proposed to automate seizure type classification (STC). However, ML models especially deep networks (DNs) learn from datasets randomly supplied in mini-batches leading to overfitting and poor generalization. Additionally, computational demands for DNs especially a convolutional neural network (CNN) can be overwhelming. Therefore, a curriculum learning (CL) framework has been proposed in this study to solve the abovementioned challenges. CL is an incremental learning process where the model learns to generalize better by increasing the difficulty of learning tasks. Hence, we have divided the publicly available Temple University Hospital (TUH) dataset into various levels of difficulty. We aim to have a computationally efficient model capable of generalizing to unseen datasets. Therefore, we schedule binary classifications for easy tasks, ternary classifications for medium tasks, and multiclass for a hard task of STC. By training the CNN using the proposed CL, we achieved improved performance with precision, recall, F1-score, and accuracy of 84.94%, 80.29%, 82.33%, and 80.29% respectively, with an improvement of 2.02%, 2.65%, 2.58%, and 2.65% when compared with the traditional training method. Moreover, the implementation of CL allows the training time to be reduced by 170.88 s.