Mathematical Models for Seizure Source Localization in Neonates Using Machine Learning and Finite Element Method

Neonatal convulsions are one of the most common emergency neurological events in the early period after birth with the frequency of 1.5 to 3 in 1000 live births. Consequently, monitoring of the neonatal brain activity is a standard procedure implemented in neonatal intensive care units (NICUs). As a result, the amount of data generated by continuous monitoring during the infant stay at NICU is rather large and thus impossible to be completely analyzed/overviewed by pediatric neurologist. To this purpose various automated systems for seizure detection have been proposed as these seizures are important to be adequately monitored in order to attempt to remedy them and/or reduce detrimental effects this condition may have on the development of a patient. Electroencephalography is a commonly used technique to detect temporal changes and detect these seizures but lacks desired spatial resolution. To this purpose advanced signal processing algorithms are needed but their accuracy often relies on adequate geometry information which is often missing as neonatal patients are rarely subjected to high energy image acquisition. In this paper we propose a source localization algorithm that uses machine learning and inverse finite-element electromagnetic (EM) models that have potential of estimating seizure locations without the need for obtaining accurate geometry information for every patient. As a preliminary approach we evaluate the proposed algorithm using computer simulated large data set using simplified geometry of spherically shaped neonatal head.