Automatic Detection of Epileptic Spikes in Intracerebral EEG with Convolutional Kernel Density Estimation

Analyzing the electroencephalographic (EEG) signal of epileptic patients as part of their diagnosis is a very long and tedious operation. The most common technique used by medical teams is to visualize the raw signal in order to find pathological events such as interictal epileptic spikes (IESs) or abnormal oscillations. More and more efforts are being adopted to try to facilitate the work of doctors by automating this process. Our goal was to analyze signal density fields to improve the visualization and automatic detection of pathological events. We transformed the EEG signal into images on which we applied a convolution filter based on a Kernel Density Estimation (KDE). This method that we propose to call CKDE for Convolutional Kernel Density Estimation allowed the emergence of local density fields leading to a better visualization as well as automatic detection of IESs. Future work will be necessary to make this technique more efficient, but preliminary results are very encouraging and show a high performance compared to a visual inspection of the data or some other automatic detection techniques.