Abstracts

Rationale: The incidence of seizures in the neonatal period is higher than at any other point in life. As recognition is based on often subtle stereotypical movement observation, improving identification in real time for prompt treatment initiation is crucial. Currently for in-patient settings, the gold standard of long-term video EEG typically necessitates technician teams with physician supervision/review. Time-consuming and error-prone, seizures are often missed altogether. Substantial barriers exist for improvement of seizure detection: (1) lack of budget for newer, advanced systems for NICUs, and (2) lack of trained neonatal specialists to collect, monitor and review video EEG. An innovative solution is artificial intelligence to identify seizures in EEG data and deploy an automated tool to fully develop a neonatal neurointensive care program. Our previous research has obtained high accuracy using machine learning (ML) to identify seizures, but the barrier of NICU staffing to place EEG leads still exists. Here, we test the ability of ML models to identify seizures in neonates using reduced montages.

Methods: To determine how the number of EEG channels impacts the efficacy of automated seizure detection, each ML model was initially trained using data from a reputable public online data repository (Stevenson et al., 2019) containing EEG from 73 NICU patients. The EEG was collected using the International 10-20 system with a 19-channel montage and sampled at 256 Hz, digitally band-pass filtered from 0.5 to 40 Hz offline, and resampled at 128 Hz. Based on our prior research, seizure and non-seizure clips were then re-clipped to four second clips for classification. Three montages were tested: the full 19 channel, an 8-channel, and a 6-channel. The ML models were built in Python 3.9 and the Tensorflow-GPU platform and trained using a dual GPU (NVIDIA Quadro RTX 8000) system. The model architecture was a fully-convolutional neural network with 29,274 trainable parameters for the 19 channel montage, 18,010 for 8, and 15,962 for 6.

Results: For each montage, the model was trained for 500 epochs. Figure 1 shows both training and validation accuracy were highest for the full 19 channel montage, with a drop-off for the lower channel count montages. Further exploration of the model training shows in Figure 2 that the slope of both the training and validation accuracy were still increasing after 500 epochs and suggests additional training would improve accuracy of the models for each montage.

Conclusions: The ML models trained for each montage show even low numbers of EEG channels can aid in seizure detection. The full 19 channel montage performed best and therefore decisions about using fewer EEG electrodes in a NICU montage to reduce burden must be balanced against the loss of information. Further analysis is needed to identify an optimal montage and then the trained models tested against EEG collected across multiple NICU settings to determine the promise of ML in identifying seizures in neonates.

Funding: Not applicable