In this diagnostic study, we trained an AI model on 30,493 EEGs to separate normal from abnormal recordings, and then classify abnormal recordings into epileptiform focal, epileptiform-generalized, non-epileptiform-focal, and non-epileptiform-diffuse categories. We validated the AI model using three independent test datasets, consisting of 9,945 EEGs not used for training. The AI model achieved diagnostic accuracy similar to human experts.

​

The expert-level performance of our model warrants its application in remote and underserved areas. Its use has the potential to decrease EEG misinterpretation and circumvent the problem of low interrater agreement in many places where clinical EEG is read by physicians without fellowship training, without access to expert supervision, or with limited experience (often in the general neurology practice setting). Furthermore, the AI model may help reduce the workload in centers where experts are available but overburdened by clinical workloads that include EEG interpretation.

 

Since this AI model appears to identify normal EEGs with a 100% precision, experts may decide to spend less time evaluating these recordings and more time on some of the more difficult EMU or ICU recordings. The model is currently being integrated with one of the most widely used clinical EEG equipment systems (Natus Neuro, USA).

​

The exceptional advancements achieved through SCORE-AI have been acknowledged and published in JAMA Neurology, a prestigious medical journal renowned for its contributions to neuroscience.