DESIGN AND IMPLEMENTATION OF NEO-NATAL SEIZURE PREDICTION USING XGBOOST

Neonatal seizure is a significant neurological emergency. However, many neonatal seizures often go unnoticed because of their subtle or electrographic nature. Thus, it is important to identify neonatal seizure at an early stage to avoid any significant damage to the child’s brain. In the following paragraphs, the design and implementation of an automated neonatal seizure prediction system using an Extreme Gradient Boosting algorithm have been explained. Since there were stringent restrictions on using any clinical data for research purposes, a realistic approach to generating a neonatal EEG dataset was adopted for the study. The electroencephalogram was converted into discriminative feature vectors using time domain, frequency domain, and non-linear characteristics. The proposed XGBoost algorithm was implemented using a 5-fold cross-validation approach for evaluating its performance with respect to other machine learning algorithms such as Logistic Regression, Support Vector Machine, and Random Forest. From the experimental results, it is evident that the proposed XGBoost algorithm was able to achieve an accuracy of 93.2%, sensitivity of 92%, specificity of 94%, and area under the ROC curve of approximately 0.97.The feature importance analysis showed that frequency-based and signal complexity features were most discriminative for seizure detection. The results demonstrate that XGBoost is a strong method for seizure detection, which is computationally efficient and clinically interpretable, especially when resources are limited in neonatal ICUs. that traditional machine learning with careful feature design can achieve competitive performance with interpretability.This work shows that traditional machine learning with careful feature design can achieve competitive performance with interpretability.