Publications Repository - Gdańsk University of Technology

Preventing early progression of epilepsy and sothe severity of seizures requires effective diagnosis. Epileptictransients indicate the ability to develop seizures but humansoverlook such brief events in an electroencephalogram (EEG)what compromises patient treatment. Traditionally, trainingof the EEG event detection algorithms has relied on groundtruth labels, obtained from the consensus of the majority oflabelers. In this work, we go beyond labeler consensus on EEGdata. Our event descriptor integrates EEG signal features withone-hot encoded labeler category that is a key to improvedgeneralization performance. Notably, boosted decision treestake advantage of singly-labeled but more varied training sets.Our quantitative experiments show the proposed labeler-hotepileptic event detector consistently outperforms a consensus-trained detector and maintains confidence bounds of the de-tection. The results on our infant EEG recordings suggestdatasets can gain higher event variety faster and thus betterperformance by shifting available human effort from consensus-oriented to separate labeling when labels include both, the eventand the labeler category.

Authors

• Łukasz Czekaj,
• Wojciech Ziembla,
• Paweł Jezierski,
• Paweł Świniarski,
• Anna Kołodziejak,
• Paweł Ogniewski,
• Paweł Niedbalski,
• dr inż. Anna Węsierska link open in new tab ,
• dr inż. Daniel Węsierski link open in new tab