Abstracts

Rationale:
In our study, we aimed to contribute a practical advancement in epilepsy research with direct applicability in clinical settings. Despite numerous studies focusing on seizure prediction models, a significant portion of epileptic patients continues to face uncertainties in their condition management. In our endeavor, we not only introduced an efficient and high-performing predictive model but also delved into the nuanced interpretation of results, considering both patient background and seizure type. Our approach seeks to bridge the gap between theoretical advancements and real-world effectiveness, offering insights crucial for personalized patient care.




Methods: In our study, we introduced an epileptic prediction model built upon a two-layer Long Short-Term Memory (LSTM) architecture. Our model stands out for its commendable performance and low complexity, rendering it particularly suitable for integration into wearable and implantable devices. This feat was achieved by judiciously leveraging just four electrodes, comprising a minimum of two origin electrodes, thereby significantly simplifying the model's architecture.

Results: To evaluate our model, we conducted experiments using data from twenty-six patients sourced from the European iEEG Dataset, with an average AUC of 0.885. Additionally, we conducted a detailed investigation into the influence of seizure type on the predictive outcomes. Our findings revealed noteworthy disparities in sensitivity across different seizure categories. Specifically, we observed that the reported sensitivity for unclassified (U) and simple partial (SP) seizure categories tended to be lower compared to that of complex partial (CP) and secondary generalized (SG) seizure types. This disparity suggests that our epileptic prediction system may exhibit superior performance in patients experiencing CP and SG seizures relative to those presenting with simple partial and unclassified seizures. Consequently, discerning the seizure type becomes pivotal in interpreting the efficacy of the predictive system for each patient. We also investigated the patient-specific details such as age and gender, seizure types, and surgery results on the
system’s performance. The results indicated that however these valuables are not directly considered in making the model,
they can have an effect on the results since they affected the data during the recordings. Understanding how these factors
impact the efficacy of prediction systems can lead to more personalized and effective interventions for individuals with
epilepsy.


Conclusions: We have proposed an epileptic seizure prediction system based on a two-layer LSTM and have conducted a detailed
study, which is the first to utilize the large iEEG European dataset and analyze the impact of seizure type.
The simulation results indicate that the model with its simplest structure in conjunction with the mean method for postprocessing
achieves the best performance, with an average AUC of 0.885. The system performance is comparable to the
best of prior work despite its small model size. The results also
show that the system performance is impacted by the seizure
type and patient's characteristics.

Funding: NA