A Multifractal-Guided Machine Learning Framework for Late Post-Traumatic Seizure Prediction Following Hemorrhagic Traumatic Brain Injury

Abstract Traumatic brain injury can lead to post-traumatic epilepsy, yet early, reliable biomarkers to predict its emergence remain elusive. By investigating the multifractal characteristics of electroencephalogramrecordings from the first available day post-injury, we develop for the first time a machine learning framework that distinguishes between traumatic brain injury patients who develop late post-traumatic seizures and those who do not. Statistical analysis demonstrates statistically significant differences in multifractal properties of EEG signals between patients who develop late post-traumatic seizures and patients who do not. We show that random forest classifier trained on multi-fractal properties of EEG achieve a high predictive accuracy (95%) and area under the curve (98%) for predicting late PTS. The predictive power of multifractal features was robust to sample length and electrode selection. Our findings indicate that multifractal properties of EEG offers a promising, objective approach to early risk stratification for post-traumatic epilepsy in neurocritical care settings.