Paper ID
2536
Paper Type
full
Description
We applied machine learning to diagnose epilepsy based on the fine-graded spectral analysis of seizure-free (resting state) EEG recordings. Despite using unspecific agglomerated EEG spectra, our fine-graded spectral analysis specifically identified the two EEG resting state sub-bands differentiating healthy people from epileptics (1.5-2 Hz and 11-12.5 Hz). The rigorous evaluation of completely unseen data of 100 EEG recordings (50 belonging to epileptics and the other 50 to healthy people) shows that the approach works successfully, achieving an outstanding accuracy of 99 percent, which significantly outperforms the current benchmark of 70% to 95% by a panel of up to three experienced neurologists. Our epilepsy diagnosis classifier can be implemented in modern EEG analysis devices, especially in intensive care units where early diagnosis and appropriate treatment are decisive in life and death scenarios and where physicians’ error rates are particularly high. Our approach is accurate, robust, fast, and cost-efficient and substantially contributes to Information Systems research in healthcare. The approach is also of high practical and theoretical relevance.
Recommended Citation
Buettner, Ricardo; Frick, Janek; and Rieg, Thilo, "High-performance detection of epilepsy in seizure-free EEG recordings: A novel machine learning approach using very specific epileptic EEG sub-bands" (2019). ICIS 2019 Proceedings. 16.
https://aisel.aisnet.org/icis2019/is_health/is_health/16
High-performance detection of epilepsy in seizure-free EEG recordings: A novel machine learning approach using very specific epileptic EEG sub-bands
We applied machine learning to diagnose epilepsy based on the fine-graded spectral analysis of seizure-free (resting state) EEG recordings. Despite using unspecific agglomerated EEG spectra, our fine-graded spectral analysis specifically identified the two EEG resting state sub-bands differentiating healthy people from epileptics (1.5-2 Hz and 11-12.5 Hz). The rigorous evaluation of completely unseen data of 100 EEG recordings (50 belonging to epileptics and the other 50 to healthy people) shows that the approach works successfully, achieving an outstanding accuracy of 99 percent, which significantly outperforms the current benchmark of 70% to 95% by a panel of up to three experienced neurologists. Our epilepsy diagnosis classifier can be implemented in modern EEG analysis devices, especially in intensive care units where early diagnosis and appropriate treatment are decisive in life and death scenarios and where physicians’ error rates are particularly high. Our approach is accurate, robust, fast, and cost-efficient and substantially contributes to Information Systems research in healthcare. The approach is also of high practical and theoretical relevance.