Inimeste intrakraniaalsete EEG salvestiste põhjal epileptiliste hoogude ennustamiseks sobivate tunnuste analüüs
Files
Date
2017
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Epilepsiahooge on üritatud ennustada mitmeid aastakümneid, kasutades tipptasemel tunnuseid ja masinõppemeetodeid. Kui õnnestuks välja töötada süsteem, mis reaalajas hoiatab patsiente eelseisvate hoogude eest, parandaks see oluliselt patsientide elukvaliteeti. Epilepsiahoogude ennustamine koosneb kahest etapist: tunnuste ekstraheerimine ning näidiste klassifitseerimine hoogudevaheliseks (tavaline ajaperiood) või hooeelseks. Enamasti kasutatakse EEG andmeid, sest EEG on odav, transporditav ning väljendab muutusi ajudünaamikas kõige täpsemini. Kui enamik uuringuid keskendub uudsete tunnuste ekstraheerimisele või uute klassifitseerimisalgoritmide rakendamisele, siis antud bakalaureusetöö eesmärk oli välja selgitada, missugused kasutatavad tunnused on kõige olulisemad. Kui on teada, missugused tunnused kõige rohkem mõjutavad ennustamistulemusi, aitab see paremini aru saada nii klassifitseerimisalgoritmide tööprotsessist kui ka ajudünaamikast ning vähendada tunnuste hulka, mida masinõppes kasutada, muutes seega klassifitseerimisprotsessi efektiivsemaks. Bakalaureusetöös kasutati kahe patsiendi intrakraniaalseid EEG andmeid ning kolme algoritmi scikit-learn teegist, mida kombineeriti meetoditega, mis hindavad tunnuste mõju. Saadud ennustustäpsused olid mõõdukalt head kuni suurepärased ning võimaldasid seega analüüsida tunnuste mõju usaldusväärselt iga klassifitseerimisalgoritmi kohta.
Epilepsy seizure prediction is a challenge that scientists have tried to overcome throughout many decades, using different state-of-the-art features and machine learning methods. If a forecasting system could predict and warn epilepsy patients of impeding seizures in real time, it would greatly improve their quality of life. Seizure prediction consists of two stages: feature extraction from the data and sample classification to interictal (non-seizure) or preictal (preseizure) state. EEG data is commonly used, as it is inexpensive, portable and it most clearly reflects the changes in the brain’s dynamics. While most studies focus on extracting novel features or using new classifiers, this Thesis focuses on ascertaining the most significant features among some that are commonly used in seizure prediction. Knowing which features influence the prediction results the most, helps to understand the inner workings of both the classifiers and the brain activity and to reduce the feature set in future research, making the classification process more effective. Intracranial EEG data of two patients was used in this Thesis with three classifiers from the scikit-learn library, which were combined with methods for evaluating feature importance. Moderately good to excellent prediction accuracies were achieved with these methods, which allowed to reliably analyze the feature importance results of the different classifiers.
Epilepsy seizure prediction is a challenge that scientists have tried to overcome throughout many decades, using different state-of-the-art features and machine learning methods. If a forecasting system could predict and warn epilepsy patients of impeding seizures in real time, it would greatly improve their quality of life. Seizure prediction consists of two stages: feature extraction from the data and sample classification to interictal (non-seizure) or preictal (preseizure) state. EEG data is commonly used, as it is inexpensive, portable and it most clearly reflects the changes in the brain’s dynamics. While most studies focus on extracting novel features or using new classifiers, this Thesis focuses on ascertaining the most significant features among some that are commonly used in seizure prediction. Knowing which features influence the prediction results the most, helps to understand the inner workings of both the classifiers and the brain activity and to reduce the feature set in future research, making the classification process more effective. Intracranial EEG data of two patients was used in this Thesis with three classifiers from the scikit-learn library, which were combined with methods for evaluating feature importance. Moderately good to excellent prediction accuracies were achieved with these methods, which allowed to reliably analyze the feature importance results of the different classifiers.