Abstracts

This is a Late-Breaking abstract.

Rationale: The identification of epileptic seizures from electroencephalograms (EEGs) is a tedious process which requires a trained eye and numerous man hours. The development of a reliable, automated, and patient-independent seizure detector is thus crucial in providing timely assessment of EEGs and helping neurologists focus on providing care to the patients.

Methods: To this end, we propose a machine-learning based patient-independent automatic seizure detector to detect seizures in both scalp EEGs (sEEG) and intracranial EEGs (iEEG).The seizure detector that we have developed performs seizure detection in stages. The detector first identifies seizures in each EEG channel (channel-level detection). This information is then used to detect seizures at the segment-level (segment-level detection), and to produce the start and end point of seizures (EEG-level detection). At the channel-level, we trained a convolutional neural network (CNN) and transformers (TRF) model with belief matching (BM) to predict the probability of a seizure in each EEG channel. Regional statistical features of the channel-level seizure probabilities for each segment are extracted and used as input to a XGBoost model, which outputs the probability of a seizure at the segment-level. Finally, postprocessing procedures are applied to the segment-level seizure probabilities to obtain the start and end points of the seizures. To evaluate the performance of our detector, the minimum overlap evaluation scoring (MOES) metric is developed. MOES is more stringent than existing metrics as it considers the percentage overlap between detected and annotated seizures to reward detections whose start and end times match the annotations more closely. The models are trained on the Temple University Hospital Seizure (TUSZ) sEEG dataset and evaluated on 6 independent datasets.

Results: The proposed detector achieves a sensitivity (SEN) of 0.617-1.000, precision (PRE) of 0.534-1.000, average false positive rate per hour (aFPR/h) of 0.425-2.002, median false positive rate per hour (mFPR/h) of 0.000-1.003 for all adult EEGs across four datasets. Meanwhile, the detector achieves SEN of 0.227-0.678, PRE of 0.377-0.818, aFPR/h of 0.253-0.421 and mFPR/h of 0.118-0.223 for all neonatal and paediatric EEGs across two datasets. When benchmarked against Persyst 14 on a proprietary dataset, our detector is shown to have lower false positive rate and higher sensitivity, especially towards focal seizures, as well as better segmentation between distinct seizures.

Conclusions: The proposed detector, which performs seizure detection in stages starting from individual channels, also performs significantly better compared to  seizure detection algorithms in the literature that directly process all channels simultaneously, without processing each channel individually. Such models tend to be much more complex than the proposed approach and generalize much worse to independent datasets, as also shown by our results. _x000D_  
Funding: This study is funded in part by the National Health Innovation Centre (NHIC) in Singapore.