212 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 54, NO. 2, FEBRUARY 2007 Epileptic Seizure Detection Using Genetically Programmed Artificial Features Hiram Firpi*, Member, IEEE, Erik D. Goodman, and Javier Echauz, Member, IEEE Abstract—Patient-specific epilepsy seizure detectors were designed based on the genetic programming artificial features algorithm, a general-purpose, methodic algorithm comprised by a genetic programming module and a -nearest neighbor classifier to create synthetic features. Artificial features are an extension to conventional features, characterized by being computer-coded and may not have a known physical meaning. In this paper, arti- ficial features are constructed from the reconstructed state-space trajectories of the intracranial EEG signals intended to reveal patterns indicative of epileptic seizure onset. The algorithm was evaluated in seven patients and validation experiments were carried out using 730.6 hr of EEG recordings. The results with the artificial features compare favorably with previous benchmark work that used a handcrafted feature. Among other results, 88 out of 92 seizures were detected yielding a low false negative rate of 4.35%. Index Terms—Epilepsy, feature extraction, genetic program- ming, seizure detection, state-space reconstruction. I. INTRODUCTION A PPROXIMATELY one percent of the world is affected by epilepsy, a bewildering neurological disorder that may cause brief electrical disturbances in the brain, producing a change in the sensation, awareness, and behavior and which is typified by recurrent seizures. Twenty-five percent of these cases cannot be fully controlled using current therapies, medi- cations, or surgical treatments. For these people, a stimulation device would be of great benefic. From a generalistic point of view, an implantable device would be compounded by a detector component and a control/stimulation component. In this paper, our efforts are devoted to present a new tool for the detection of epileptic seizures. The introduction of the electroencephalogram (EEG) has brought cues characterizing the dynamics of the epilepsy. Although detection of epileptic seizures is easier than the prediction problem, is still far from trivial. There are many anomalies that occur naturally in the EEG signals that trigger the detectors and declare that a seizure is occurring when actually Manuscript received July 31, 2005; revised June 24, 2006. Asterisk indicates corresponding author. *H. Firpi is with the Center for Computational Biology & Bioinformatics, Indiana University-Perdue University, 410 W. 10th Street, Suite 5000, Indi- anapolis, IN 46202 USA (e-mail: hfirpi@ieee.org). E. D. Goodman is with the Department of Electrical and Computer Engi- neering, Michigan State University, East Lansing, MI 48824-1226 USA (e-mail: goodman@egr.msu.edu). J. Echauz is with the BioQuantix Corp., Atlanta, GA 30363 USA (e-mail: echauz@ieee.org). Color versions of Figs. 2 and 5–7 are available online at http://ieeexplore. ieee.org. Digital Object Identifier 10.1109/TBME.2006.886936 it is not. False starts (i.e., high-frequency, low amplitude events similar to those occurring at the beginning of a seizure), delta trains, and spike-and-wave discharges, all of them lead to false alarms and, thus, would medicate the patient unnecessarily. Additionally, seizures must be detected as soon as possible so the control/stimulation can be delivered immediately and control the seizures without further consequences. Therefore, there is a tradeoff between the number of false alarms and the number of false negatives (when the detector says a seizure is going to occur and it does not), and consequently affecting how early a seizure can be detected. Several algorithms have been proposed to detect epileptic seizure onsets [1], [5], [6], [10], [15], [18], [20], [21], [24], among others in the search for an accurate detector. Some ap- proaches that apply digital signal processing or filter theory have acceptable performance. These algorithms rely on one or more features—informative measures or attributes eluded from raw data—to decide whether a seizure is occurring or not. However, to extract the relevant information that can facilitate such detec- tion, features are calculated using conventional techniques and methodologies that are time-consuming, tedious, and trial-and- error processes requiring a great deal of effort from researchers. All of these conventional techniques rely on gathering a feature or a set of features conceived by knowledge of a feature for- mula or algorithm that may have been obtained from intuition, tradition, the physics of the problem, or analogies to problems in other fields. Because of this, we wonder if there is any under- lying pattern or patterns that are being ignored by these tradi- tional methods. The relevancy provided by the features is lim- ited by the attributes measured for such features. In pursuing a methodology that surpasses the aforementioned limitations of the conventional features, we develop an algo- rithm that systematically and automatically can find or generate artificial features or attributes starting from raw data—in this case, from the IEEG signals. Artificial features are defined by features that are computer-designed, learned, inductive, opti- mized, and data-driven. These are desirable properties on fea- tures because they are meant to pay attention, capture, extract, or uncover any underlying, useful pattern that might be suppressed by the conventional features. We will see that unlike most ap- proaches, this work is not based solely on tuning a few param- eters of fixed terms to find the “best” detector for a patient, but we are using an algorithm that also changes the structure of the equations, giving us enough flexibility to design an “optimal” (or at least highly tuned) feature that is sensitive to the charac- teristics or patterns of a given patient. Section II explains the methodology on how to reconstruct the state-space trajectory from the IEEG signal. It also explains the 0018-9294/$25.00 © 2007 IEEE