Abstracts

Rationale: Machine learning algorithms for detecting epileptic abnormalities are improving to the point of matching expert human performance. Interictal epileptiform discharges (IEDs) convey clinically meaningful information and are evaluated for clinical trial endpoints. Current algorithms to detect IEDs in EEG are not necessarily robust to missing electrodes, a common real-world data scenario. Such robustness is of critical importance, as it is common that human experts are able to identify IEDs even in the presence of one or more missing channels. In addition, novel EEG solutions utilizing reduced channel counts are emerging. Here we demonstrate the performance of an IED detection algorithm that matches human expert agreement to understand robustness to missing channels.

Methods: Eight human experts labeled IEDs in 1063 10-20 EEG recordings from 1051 patients. These patients ranged in age from hours to 100 years old, were 51% female, and included individuals with and without neurological illness. A deep neural network was trained on this data to estimate whether a panel of human experts would determine that an IED was present in 1-second EEG segments. Model performance was assessed after removing 0 to 18 channels from a held-out dataset. Channel removal was performed along longitudinal or transverse bipolar chains starting anteriorly or posteriorly.

Results: The algorithm demonstrated near ideal calibration to consensus human opinion, and resulted in an AUC of 0.99 on high-agreement observations (≥7 experts agree) (Figure 1). With sequential channel removal, the algorithm maintained reasonable performance even to large numbers of deleted channels (Figure 2). Channel removal resulted in the worst performance when left frontotemporal electrodes were deleted first and best performance was observed with early loss of occipital leads. AUC remained above 0.75 even when only 4 channels remained, as long as these included temporal channels. Performance was substantially higher when restricting to high-agreement observations (≥7 experts agreed), with >0.75 AUC for all channel deletions except those involving the entire left temporal chain.

Conclusions: IED detection is feasible using deep learning networks, can be robust to electrode loss, and a single algorithm may be capable of handling a wide variety of electrode configurations. This form of quantification allows for more accurate EEG analytics that are impossible for human reviewers (such as real-time IED alarms, accurate characterization of IED burden/changes over long recordings, and IED topography). This analysis is limited by the use of unselected IED types (including generalized IEDs) and arbitrary channel removal order, which will be addressed in future work. To our knowledge, this work represents the first systematic demonstration of a machine learning-based epileptiform discharge detector that is robust to real-world scenarios where individual channels may be missing. As a result, we are able to harness IED-derived metrics as robust biomarkers for clinical trials and neurodiagnostics.

Funding: This work was supported by Beacon Biosignals.