Towards an Online Seizure Advisory System—An Adaptive Seizure Prediction Framework Using Active Learning Heuristics

• Main Authors: Vignesh Raja Karuppiah Ramachandran, Huibert J. Alblas, Duc V. Le, Nirvana Meratnia
• Format: Article
• Language: English
• Published: MDPI AG 2018-05-01
• Series: Sensors
• Subjects: implantable body sensor networks; health-care; EEG; epilepsy; signal processing; machine learning; seizure prediction
• Online Access: http://www.mdpi.com/1424-8220/18/6/1698

In the last decade, seizure prediction systems have gained a lot of attention because of their enormous potential to largely improve the quality-of-life of the epileptic patients. The accuracy of the prediction algorithms to detect seizure in real-world applications is largely limited because the brain signals are inherently uncertain and affected by various factors, such as environment, age, drug intake, etc., in addition to the internal artefacts that occur during the process of recording the brain signals. To deal with such ambiguity, researchers transitionally use active learning, which selects the ambiguous data to be annotated by an expert and updates the classification model dynamically. However, selecting the particular data from a pool of large ambiguous datasets to be labelled by an expert is still a challenging problem. In this paper, we propose an active learning-based prediction framework that aims to improve the accuracy of the prediction with a minimum number of labelled data. The core technique of our framework is employing the Bernoulli-Gaussian Mixture model (BGMM) to determine the feature samples that have the most ambiguity to be annotated by an expert. By doing so, our approach facilitates expert intervention as well as increasing medical reliability. We evaluate seven different classifiers in terms of the classification time and memory required. An active learning framework built on top of the best performing classifier is evaluated in terms of required annotation effort to achieve a high level of prediction accuracy. The results show that our approach can achieve the same accuracy as a Support Vector Machine (SVM) classifier using only 20 % of the labelled data and also improve the prediction accuracy even under the noisy condition.