In this study, we have undertaken an exhaustive comparison of common and SOZ related hybrid iEEG markers across a range of machine learning models, with a specific focus on their effectiveness in predicting successful epilepsy surgical outcomes. Our findings reveals that the GNN, underpinned by hybrid markers, achieves an impressive accuracy of over 94%, surpassing other models by more than 6.59% (p = 0.0412). This remarkable accuracy is achieved using the GNN model paired with hybrid markers, whether in singular or combined feature. In the single marker prediction, the spectral feature of high-gamma band has great performance, which is consistent with the current mainstream concepts such as the influence of high-frequency oscillation on epilepsy, suggesting that high-frequency band (> 90 Hz) cannot be ignored in the study of epilepsy origin6–8.
Epilepsy surgery is crucial for treating patients with drug-resistant epilepsy, and improving the success rate of surgery is of utmost importance. Even a small increase in prediction accuracy would represent a significant step forward. Numerous researchers have proposed various SOZ localization markers and epilepsy localization models to facilitate the assessment of the feasibility and efficacy of surgical plans by clinicians before surgery. Some of them also utilize machine learning algorithms to predict surgical outcomes, achieving accuracy rates ranging from 60–95%.20,22,24 Our accuracy surpasses the results reported in the majority of these research studies. Particularly noteworthy is that studies employing hybrid label-driven machine learning algorithms consistently demonstrated strong performance on their own datasets, affirming the viability and effectiveness of this approach.23,29 Our study presents a comprehensive "combination boxing" approach with an impressive prediction accuracy of 94.30%. With the incorporation of additional data, markers, or machine learning algorithms, we can further enhance the precision of our predictions using this method.
Machine learning is an effective tool for recognizing data patterns with unclear mechanisms. However, training an unimproved machine learning model may yield inadequate results. In this study, we compare the performance of a common neural network with that of a GNN. The GNN outperformed in both single-marker and hybrid-marker scenarios. In particular, the GNN achieve an accuracy of 75.68% for a single marker and 94.30% for hybrid markers, whereas the common neural network achieved an accuracy of only 73.18% for a single marker and 85.70% for hybrid markers. Notably, the author of the fragility work use deformation models of traditional random forests to predict surgical outcomes, whereas we propose a genetic algorithm-based neural network that can expand the range of machine learning models with efficient prediction ability.30
Furthermore, we discover certain counterfactual scenarios when we use this approach to review and analyze previous surgical resection plans.
In Fig. 6, the left picture indicates the patient’s original EEG during the seizure, whereas the table on the right indicates the SOZ seizure areas identified by the seven markers. The surgical column shows the channels in the surgical area (indicated by the red font in the highlighted part of the table). Patient pt13 exhibit successful surgical outcomes, and our GNN model with hybrid markers predict the success of the surgery. The predictions of the seven markers closely matched the resection expectations of clinicians, indicating that the markers could simulate clinician judgment. Similarly, the surgery for patient 60 fails, and our model accurately predict surgical failure. We observed that the predictions of the markers are strikingly incongruous with the resection area of clinicians, suggesting that our model exhibited better judgment than the clinicians in some cases. For patient pt14, the predictions of markers matched the clinical viewpoints, and our model predicted successful surgery every time. Unfortunately, the surgery failed, suggesting that the failure is not attributed to inaccurate positioning. This failure may have been because of other factors such as postoperative rehabilitation and the formation of new lesions.2,5 In summary, our GNN model with hybrid markers demonstrate SOZ localization abilities that are comparable to those of clinicians and provided accurate predictions of surgical results
Currently, the inadequate availability of high-quality data and incomplete surgical information are major obstacles to accurately predicting epilepsy surgery outcomes. The scarcity of data makes it challenging to apply large-scale models and use high-dimensional inputs, which in turn complicates the validation of the reliability of each SOZ marker. Insufficient data not only relates to the current state of surgical treatment but also presents ethical considerations in medical practice. Furthermore, SOZ markers from different perspectives rely on their assumptions and unpublished data, complicating the replication of previous studies and making it difficult to evaluate the reliability of these markers. Moving forward, we aim to gather more high-quality, multicenter data to assess several SOZ markers to improve the surgical management of patients with drug-resistant epilepsy.