Classification of epileptic motor manifestations using inertial and magnetic sensors

Abstract

In order to characterize objectively the succession of movements observed during motor seizures, inertial and magnetic sensors were placed on epileptic patients. Video recordings synchronized with motion recordings were analyzed visually during seizures and divided, for each limb, into events corresponding to different classes of motor manifestations. For each classified event, features were extracted and a subset selection was automated using artificial neural networks. The best artificial neural network was simulated on whole recordings to generate a stereotypic evolution of motor manifestations that we called motorograms. It is shown that motorograms can point out seizure movements and emphasize epileptic patterns.

Introduction

Epileptic seizures are manifestations of brain dysfunction, due to hypersynchronous discharges either located in part of the cortex referred to as partial seizures or involving the two hemispheres referred to as generalized seizures. The prevalence of epilepsy is important since 0.5 to 1% of the average population is concerned, with an important social and economic impact [1]. There are several types of epilepsies and efforts are made to classify them into similar syndromes [2]. Also there exist many seizure types, of which a significant number are characterized by motor symptoms. In 70% of the cases, seizures are controlled by using antiepileptic drugs. The remaining 30% are drug-resistant patients, and for those among them whose seizures are focal corresponding to $\approx 60\%$ of the cases, surgery can be discussed in about one fourth of the cases, which represents about 10,000 patients in France. Surgery consists in the removal, disconnection or coagulation of the brain areas from where seizures arise, the epileptogenic zone, in order to render the patients seizure-free. Such an approach requires to precisely localize the epileptogenic zone. A number of exams are used, among which seizure recording is of crucial importance. Such recordings are performed in an epilepsy monitoring unit (EMU), using an electroencephalographic (EEG) system coupled to a video system (video-EEG). The clinical analysis of epileptic seizures is of particular importance to localize the epileptogenic area, but video analysis remains essentially qualitative. It is therefore subjective, explaining why the precise characterization of epileptic motor manifestation and their localizing significance remains a matter of debate.

Few solutions have been proposed to quantify movements during epileptic seizures and to provide a more objective evaluation of epileptic motor symptoms:

• Some authors use specialized video processing methods with or without markers set on patient's limbs [5], [3], [4]. The major benefit of such approaches is that they rely on a device already available at the patient's bedside. One drawback is the poor resolution of the quantification of movement and the difficulty to reconstruct the movement from 2D images. Another shortcoming is when some markers disappear from the camera field of view, leading to uncertainties.

• Another equipment used in medicine for the quantification of movement is the accelerometer. In neurology, this sensor has been studied principally for Parkinson's disease [6], [7], [8], [9] and the detection of hand tremors [10]. 3D accelerometers have also been used for epilepsy [11], [12] where visual analysis of signals revealed stereotypical patterns of motor seizures and gave clues for automatic analyses. The advantage of accelerometer sensors is their low cost and their low energy consumption enabling ambulatory monitoring. The drawback is an indetermination of a class of movement in the horizontal plane, like undesired rotation of the head to the right or left, frequently observed during versive seizures. In order to obtain a full characterization of 3D movements, it is possible to couple magnetic sensors to accelerometers [13]. These will be referred to as inertial and magnetic sensors (IMS) in the remainder of the paper. This solution is cheap and with low power consumption compared to gyroscopic sensors measuring angular rates and also enabling 3D movements characterization.

The aim of our study was to investigate the use of IMS for the quantification of movements observed during motor seizures recorded as part of presurgical evaluation of patient with drug-resistant partial seizures. This quantification of movement lead us to design a classification system for the analysis of epileptic motor manifestations that we present in this article after positive results obtained on an exploratory analysis presented in [14]. For the evaluation of inertial and magnetic sensors, we positioned three of them on the limbs of epileptic patients and collected data in an EMU. For the quantification of movements, angular measurements were computed and features extracted. For the design of the classification system, signals from recorded seizures were analyzed and split into different events corresponding to classes of epileptic motor manifestations. The association of the extracted features to these classes was then realized using supervised learning. The different procedures retained for data acquisition, for feature extraction and for supervised learning are presented and empirical performances of the system are evaluated. Results of the classification system are presented using a new representation of the evolution of motor manifestations on each limb that we called motorogram.

The aim of this study was not to realize an epileptic seizure detector but was to give an objective quantification and representation of movement disorders. Besides, the discrimination of non-epileptic manifestations versus epileptic manifestations seems hard, especially with one sensor. Indeed a lot of usual manifestations can mimic locally and temporally some epileptic manifestations such as positioning to grasp something, writing, coughing and so on. We hypothesize, in this study, that whole patterns generated by classifiers on all sensors can be specific of some seizures. The motorograms were introduced to present these patterns and help the neurophysiologist in his decision making process.

First of all we will present the material used to monitor epileptic patients and to collect data. Then we will focus on methods to: split motor manifestation observed during seizures into classified events; extract features from these events to create a database; train classifiers using supervised learning techniques based on artificial neural networks. Results will present a description of signals by class, the analysis of events duration, the model obtained by supervised learning and the simulations on whole recordings to generate motorograms. The discussion will point out the benefits and the drawbacks of inertial and magnetic sensors and the methods used in the study to put in evidence epileptic motor manifestations.