Accurately foreseeing the likelihood of seizure recurrence holds paramount importance in effectively diagnosing and managing epilepsy. The foundational element in assessing the risk of seizure recurrence lies in routine electroencephalography (EEG). However, the interpretation of EEG results rests heavily on the visual detection of interictal epileptiform discharges (IEDs) by neurologists, which has its limitations in terms of sensitivity.
Unlocking the potential of automated EEG analysis could substantially enhance its diagnostic efficacy and accessibility. Our primary aim was to construct a forward-looking predictive model utilizing automated EEG processing, geared towards forecasting one-year seizure recurrence in individuals undergoing routine EEG examinations.
In a comprehensive approach, we carefully selected a consecutive group of 517 patients who had undergone routine EEG assessments at our institution (training set), and a separate cohort of 261 patients (testing set) from a different time period. Through the development of a sophisticated automated processing pipeline, we harnessed both linear and non-linear features extracted from the EEG data.
Our pioneering effort involved training advanced machine learning algorithms using multi-channel EEG segments to anticipate one-year seizure recurrence. This comprehensive approach also allowed us to scrutinize the influence of IEDs and clinical variables on the predictive performance. To validate our model, we rigorously assessed its effectiveness using the independent testing set.
The area under the receiver operating characteristic curve (ROC-AUC) for forecasting seizure recurrence one year post-EEG in the testing set reached 0.63, with a 95% confidence interval of 0.55–0.71. Even in EEG results without discernible IEDs, our predictions remained significantly above chance.
This groundbreaking discovery underscores the presence of additional EEG signal changes beyond IEDs, encapsulating the intricate dynamics of seizure propensity. Our findings pave the way for a deeper understanding of epilepsy and offer insights into avenues for improved predictive
Source Nature
