ANOLUF: A Feature Selection and Channel Selection Methodology for Medical Event-Detection

Abstract

A feature selection and a channel selection methodology are presented in this paper to provide a reliable and flexible comparison between different features. Using the proposed method, the best feature extractor(s), as well as the best corresponding channel(s) for a medical event-detector are selected. In this paper, we focus on the seizure onset detection (SOD) in closed-loop stimulation systems for epilepsy control, as the target application. However, the target application can be any type of medical event-detector and is not limited to SODs. Detecting the onset of seizures with a minimum delay and maximum sensitivity and specificity (or minimum false alarm rate) is the major and the most important task of an implantable medical device (IMD) for epilepsy control. However, delay in seizure onset detection is not used to score models or features employed in the state-of-the-art SODs. Thanks to the proposed method, feature scoring is performed based on any set of target performance metrics such as detection delay (DD), false-positive rate (FPR), false-negative rate (FNR), sensitivity, power consumption, etc. The presented method is applied to the long-term dataset of the open-access SWECETHZ iEEG database. Ten top-score features are extracted from each patient's iEEG which include the feature extractor and the corresponding electrode number. This work and its results provide a reliable methodology for designing accurate and robust event detectors for IMDs.