Abstracts

Rationale: Acute provoked neonatal seizures are associated with increased mortality and have enduring adverse impacts on neurodevelopment. Neonatal seizure forecasting may improve outcomes by enabling earlier diagnosis and treatment, as well as optimize allocation of EEG monitoring resources.

Methods: We used the Helsinki University Hospital (HUH) neonatal seizure dataset containing 79 EEGs to investigate the feasibility of forecasting neonatal seizures through machine learning of QEEG features (Stevenson et al. Sci data 6.1 2019). We evaluated varying seizure prediction horizons (SPH: 1 & 5 min) and seizure occurrence periods (SOP: 10, 20, & 30 min) on model performance (Fig 1). QEEG features previously used in seizure prediction were calculated on nonoverlapping 20 sec EEG epochs (Figure 2A). We developed a convolutional long short-term memory neural network (ConvLSTM) for next timestep prediction of preictal/interictal state and used SHapley Additive exPlanations (SHAP) analysis for QEEG feature selection into the ConvLSTM (Figures 2A-C). We defined preictal state as between 6 min to 1 min prior to seizure onset and interictal as between 30 seconds after the end of seizure to 6 min prior to the next seizure. 5 min preictal duration has previously been studied and is suitable for this dataset which has a relatively lower average interictal duration (39 min) compared to other studies (Mormann, F, et al. "On the predictability of epileptic seizures." Clin neurophysiol 116.3 2005). To address data scarcity, we used EEG data augmentation including invariant transforms and random Gaussian noise addition during ConvLSTM training. Performance was evaluated using AUROC, true positive rate of seizures detected, and false positive alarms per hour, estimated via stratified k-fold cross-validation (k=10, 68-22-10% train-val-test splits).

Results: Three EEGs were excluded due to insufficient interictal periods to permit any SOP, leaving 76 neonates (total 144 seizures) for analysis. SHAP demonstrated statistical moments, spectral power, and RQA features as robust predictors in pre-ictal and interictal classification (Figure 2A). QEEG SHAP scores lead to preictal/interictal clustering of epochs, indicating feasibility of their respective classification (Figure 2B). Sample seizure prediction output is shown in Figure 2D. The ConvLSTM predictive model achieved a 10-fold cross-validated with true positive rate of 83.3% with a false positive alarm rate of 0.41/hour, with AUROC of 0.86 with SPH of 1 min and SOP of 30 min (Figures 2E and 2F).

Conclusions: Neonatal seizure forecasting is feasible through machine learning of QEEG features.

Funding: UCSF Research and Allocation Committee (REAC) Grant