Abstracts

Rationale: Newborns have the highest risk of developing seizures. Uncontrolled seizures have a detrimental impact on the developing brain. Long-term monitoring Electroencephalography (EEG) is used to identify neonatal seizures. However, having clinicians read EEGs to identify seizures is both costly and labor-intensive. Thus, we are creating computer algorithms that use EEG data for rapid automated seizure detection.

Amplitude Integrated EEG (aEEG) is the state-of-the-art long-term monitoring algorithm for neonatal seizure detection. Despite aEEG’s widespread use, it has notable drawbacks, specifically insufficient time resolution for neonatal seizures and lack of frequency domain detail. Another monitoring tool is the Compressed Spectral Array (CSA), which displays frequency trends using spectral analysis, but is not widely used clinically for neonates. In this study, we leverage a dataset of EEG recordings from 79 patients to create an ensemble algorithm. Our goals are to improve aEEG’s performance and determine which aEEG and CSA features are most important for seizure classification.

Methods: First, we generated aEEG simulations using parameters based on two proprietary aEEG algorithms from Nicolet One and Olympic Systems. Then, we extracted time domain features from both simulations. Additionally, we increased the aEEG resolution to 15-seconds when extracting features to better resolve neonatal seizures. Next, we extracted frequency domain features using spectral analysis.
We compared the classification efficacy of each set of features using three classifiers: Random Forest (RF), Support Vector Machines (SVM), and an Artificial Neural Network (ANN). Finally, we assessed which set of features had the greatest importance for seizure classification using two feature selection methods.

Results: For seizure detection, our combination of both time and frequency features modestly outperforms individual time and frequency domain features. This trend was exemplified by the combination of features having the highest Area Under the Curve (AUC) score across all classifiers. Additionally, we found the most important features for seizure classification are: 1. the lower trace of the aEEG plot, called the Terminal Point Minimum, which was extracted from the Olympic Systems simulation, 2. the kurtosis of the aEEG extracted from the Nicolet One simulation, and 3. the power within the High Frequency Oscillation Band (80-128Hz). Finally, we compared the efficacy of only these top three features to our entire array of features. By only using the top features, we condensed our dataset to 1/7th of the original number of features. Despite downsizing, we were still able to yield AUC scores of 0.85, 0.91, and 0.96 for the ANN, SVM, and RF classifiers respectively.

Conclusions: Differences in aEEG algorithm parameters influence seizure classification. These differences can be leveraged to develop a better classifier by selecting the best features across each simulation. Additionally, High Frequency Oscillations may serve as a useful tool for seizure classification in neonates and should be further investigated.

Funding: Please list any funding that was received in support of this abstract.: The University of Chicago Medical Scientists Training Program NIGMS T32 Training Grant.