Samenvatting

Continuous brain function monitoring in a pediatric intensive care unit is often done using electroencephalography (EEG) for the detection of epileptic seizures in comatose patients. Monitoring an EEG is a time-intensive task for medical professionals, especially when the seizures manifest in a subtle way. Automatic systems that can provide support in the detection of seizures could optimize bed occupancy and improve the staffing of the nursing workforce. Moreover, deep learning methods could potentially identify biomarkers that point neurologists to new indicators of seizures.

This research is focused on how deep learning can be applied to minimally processed EEG data for the detection of epileptic seizures in pediatric comatose patients. To determine this, raw EEG data is first preprocessed by applying cleaning, constructing, and formatting techniques. Convolutional neural networks with 1D and 2D layers are trained to classify EEG fragments as seizure or non-seizure. Both these models present similar results, with a (balanced) accuracy of 0.75 and 0.76, respectively. A CNN with 1D layers acquired a higher recall of 0.83, compared to 0.81 with 2D layers. Precision is higher in a CNN with 2D layers, where it scores 0.73 compared to 0.71 in the model with 1D layers. Both models have an F1-score of 0.77. Considering the medical context, a 1D-CNN is considered to be the best performing model due to its higher recall. Although a balance between precision and recall should be achieved, minimizing the risk of missing seizures is essential.

The results lay a foundation for future research endeavors, and the recommended next step is to use data augmentation and synthesis techniques to increase the seizure data available for the improvement of model performance. More importantly, it could ensure generalizability across patients by implementing a subject-wise cross-validation approach.