Evaluating signal analysis results with machine learning algorithms: Application to electroencephalogram of epilepsy patients

Montano Hernández, María

Evaluating signal analysis results with machine learning algorithms: Application to electroencephalogram of epilepsy patients

License

Citation

Montano Hernández, M. Evaluating signal analysis results with machine learning algorithms: Application to electroencephalogram of epilepsy patients. 2017. handle: http://hdl.handle.net/10230/32898

This citation was generated automatically.

Abstract

Epilepsy is a neurological disorder that affects 50 million people worldwide. Some of these patients are pharmacoresistant and continue to suffer from seizures even under antiepileptic drugs treatment. For these patients, seizures can be potentially stopped after a resective epilepsy surgery. Using intracranial EEG in a long-term monitoring, clinical doctors localise by visual inspection what is believed to be the focal area causing the epileptic seizures. This procedure is highly observer dependent and time consuming and is carried out by highly trained neurophysiologist. Moreover, the epilepsy surgery outcome is not always satisfactory and better tools are needed to localise the focal area faster and in a more reliable way. In this work, we develop a quantitative EEG method to help in the localisation of this zone based on peri-ictal and intracranial EEG recordings. The recordings we analyse come from 6 epilepsy patients who became seizure-free after a resective epilepsy surgery. We applied a total of 12 signal analysis measures derived from linear and nonlinear signal analysis. We also used the concept of surrogates which ensures that nonlinear measures are strictly focusing on nonlinear dynamics of the signals and it has been reported to be more specific than nonlinear signal analysis measures in the localisation of the focal area. We further analysed the potential of these signal analysis measures to localise the focal area for our group of patients. Moreover, we were also interested in analysing the inter-patient variability and how this variability could affect the performance of a binary classifier. To achieve that, we explored the overlap between the focal and the nonfocal signals for each signal analysis measure and for each individual patient. Subsequently, we analysed the performance of the same signal analysis measures and the same classification task for all our database. Finally, these signal analysis measures were also used as a representation of each signal in a k-means clustering which divided the different signals into disjoint subgroups. Afterwards, we analysed to which extent these subgroup helped in the localisation of the focal area. Interestingly, we found that the usage of surrogates is not always beneficial in the focal-nonfocal classification of the signals. In concrete, when the focal signals were characterised by low frequencies, the nonlinear signal analysis measures outper- formed the surrogate corrected measures. With regard to the clustering analysis, we only found good agreement between the groups found by the algorithm and the true localisation of the focal area for 50% of our patients. In general, for these patients, there was a tendency to overdetect brain areas as focal, which is a common issue in quantitative EEG methods. The work here presented describes how 12 signal analysis measures can characterise the focal area in a heterogeneous epilepsy patients group. Therefore, we believe that this can help to understand better the dynamics in the focal area and in the healthy brain tissue