Abstracts

Rationale: Epilepsy diagnosis and management are challenged by the sporadic nature of seizures, with frequency varying from daily to yearly. Ambulatory EEG (aEEG) offers a non-invasive solution for long-term monitoring (24-48 hours or more) in such patients. However, despite its effectiveness in capturing seizures, aEEG analysis is time-consuming due to lengthy recordings and frequent artifacts, requiring meticulous review by trained epileptologists. The absence of standardized algorithms for aEEG analysis further complicates the process. This study proposes an automated seizure detection system leveraging machine learning to analyze aEEG recordings with high accuracy and specificity, augmenting the capabilities of trained epileptologists.

Methods: Electroencephalographic (EEG) recordings were acquired from 48 patients suspected of seizure activity (total duration: 3140 hours). Trained epileptologists reviewed these recordings to identify seizure events (n = 15 with seizures) which were used as the gold standard. Preprocessing involved raw data extraction & segmentation (5 sec. segments), digital filtering (1-40Hz), feature extraction (band power, signal energy, signal statistics, linelength) with z-score normalization, and subsequent labeling for machine learning (ML) algorithms. The data was split for training (70%) and testing (30%) with 10-fold cross-validation. Three distinct ML algorithms, XGBoost, CatBoost, and LightGBM, were trained on the training data and further combined (ensembled) to predict seizure occurrence.

Results: The ensemble model demonstrated high overall performance with an accuracy of 0.99, sensitivity of 0.91, specificity of 0.99, and area under the ROC curve (AUC) of 0.95. Additionally, the model exhibited a low false discovery rate (FDR), averaging 0.77 false events per hour. While individual training performance of XGBoost and LightGBM reached perfect accuracy (1.0) on all metrics, CatBoost achieved superior performance on the test dataset (accuracy: 0.99, sensitivity: 0.82, specificity: 0.99, AUC: 0.91). In the absence of established benchmarks for aEEG seizure detection, we compared our model's FDR with published results on EMU EEG datasets. The proposed method achieved a lower or equivalent FDR compared to established software (Persyst: 0.9 false events/hour, Besa: 0.7 false events/hour). Notably, while some existing software exhibits comparable FDR, it may come at the expense of missing seizures. In contrast, the proposed model achieved a sensitivity of 1.0, indicating no missed seizures as identified by trained epileptologists.

Conclusions: We have developed an ensemble machine learning model for automated seizure detection in ambulatory EEG recordings. The model achieved high overall performance and a low false positive while maintaining a perfect sensitivity (1.0), indicating no missed seizures compared to expert evaluation. The proposed model has the potential to reduce the burden on human experts by analyzing lengthy aEEG recordings while maintaining a high level of seizure detection accuracy.

Funding: The study was supported by Community Health Advocacy (C-042).