Evaluation of automatic feature detection algorithms in EEG

In this paper, we present a new method to compare and improve algorithms for feature detection in neonatal EEG. The method is based on the algorithm's ability to compute accurate statistics to predict the results of EEG visual analysis. This method is implemented inside a Java software called EEGDiag, as part of an e-health Web portal dedicated to neonatal EEG.EEGDiag encapsulates a component-based implementation of the detection algorithms called analyzers. Each analyzer is defined by a list of modules executed sequentially. As the libraries of modules are intended to be enriched by its users, we developed a process to evaluate the performance of new modules and analyzers using a database of expertized and categorized EEGs. The evaluation is based on the Davies-Bouldin index (DBI) which measures the quality of cluster separation, so that it will ease the building of classifiers on risk categories. For the first application we tested this method on the detection of interburst intervals (IBI) using a database of 394 EEG acquired on premature newborns. We have defined a class of IBI detectors based on a threshold of the standard deviation on contiguous short time windows, inspired by previous work. Then we determine which detector and what threshold values are the best regarding DBI, as well as the robustness of this choice. This method allows us to make counter-intuitive choices, such as removing the 50Hz filter (power supply) to save time. A software to improve algorithms for feature detection in neonatal EEG is proposed.The clustering quality of feature detectors with EEG rating is quantified.The method is tested on a modular definition of interburst intervals detectors.The power supply filter can be removed with little effect on the detection quality.Detectors are robust compared with the standard deviation thresholding of the EEG.