The localization of epileptic zone in pharmacoresistant focal epileptic patients is a daunting task, typically performed by medical experts through visual inspection over highly sampled neural recordings. For a finer localization of the epileptogenic areas and a deeper understanding of the pathology both the identification of pathogenical biomarkers and the automatic characterization of epileptic signals are desirable. In this work we present a data integration learning method based on multi-level representation of stereo-electroencephalography recordings and multiple kernel learning. To the best of our knowledge, this is the first attempt to tackle both aspects simultaneously, as our approach is devised to classify critical vs. non-critical recordings while detecting the most discriminative frequency bands. The learning pipeline is applied to a data set of 18 patients for a total of 2347 neural recordings analyzed by medical experts. Without any prior knowledge assumption, the data-driven method reveals the most discriminative frequency bands for the localization of epileptic areas in the high-frequency spectrum (>=80 Hz) while showing high performance metric scores (mean balanced accuracy of 0.89 +- 0.03). The promising results may represent a starting point for the automatic search of clinical biomarkers of epileptogenicity.

Multi-task multiple kernel learning reveals relevant frequency bands for critical areas localization in focal epilepsy

Vanessa D’Amario;Federico Tomasi;Veronica Tozzo;Gabriele Arnulfo;Annalisa Barla;
2018

Abstract

The localization of epileptic zone in pharmacoresistant focal epileptic patients is a daunting task, typically performed by medical experts through visual inspection over highly sampled neural recordings. For a finer localization of the epileptogenic areas and a deeper understanding of the pathology both the identification of pathogenical biomarkers and the automatic characterization of epileptic signals are desirable. In this work we present a data integration learning method based on multi-level representation of stereo-electroencephalography recordings and multiple kernel learning. To the best of our knowledge, this is the first attempt to tackle both aspects simultaneously, as our approach is devised to classify critical vs. non-critical recordings while detecting the most discriminative frequency bands. The learning pipeline is applied to a data set of 18 patients for a total of 2347 neural recordings analyzed by medical experts. Without any prior knowledge assumption, the data-driven method reveals the most discriminative frequency bands for the localization of epileptic areas in the high-frequency spectrum (>=80 Hz) while showing high performance metric scores (mean balanced accuracy of 0.89 +- 0.03). The promising results may represent a starting point for the automatic search of clinical biomarkers of epileptogenicity.
File in questo prodotto:
File Dimensione Formato  
d-amario18a.pdf

accesso aperto

Tipologia: Documento in Post-print
Dimensione 1.92 MB
Formato Adobe PDF
1.92 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11567/1015074
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
social impact