Interrater Reliability of Electroencephalogram Interpretations in Febrile Comatose African Children: Implications for...

Electroencephalography (EEG) is a diagnostic and prognostic tool used worldwide in the clinical care of comatose patients. Scalability of EEG use in resource-limited settings is constrained by multiple factors, including the lack of neurophysiologist interpreters. Although machine learning offers a path to creation of automated interpretation algorithms, development is constrained by the need for independent inputs from multiple neurophysiologists. Development may be accelerated if independent inputs have high concurrence. We assessed interrater reliability of key EEG variables from studies recorded from febrile comatose African children to better understand the need for numbers of inputs required in the development of future machine learning-driven EEG interpretation algorithms. Two independent electrophysiologists interpreted 171 electroencephalograms from children in febrile coma using standard methods. Interrater reliability was highest for the presence of electrographic seizures, with overall agreement of 96.5% and kappa of 0.781 (95% CI: 0.612-0.950). Electroencephalography variables commonly used in prognostication for cerebral malaria, a common cause of febrile coma, also had high agreement rates. Electroencephalography interpretation in Malawian children with febrile coma offers high interrater reliability in key variables used to guide clinical care and estimate prognosis. This suggests that automated interpretation methods could be developed using limited numbers of human expert trainers, accelerating development of data-driven diagnostic and prognostic algorithms for resource-limited locations. Deployment of automated EEG diagnostic and prognostic interpretation algorithms for use in high-disease-burden, low-resource settings would aid in the clinical care of febrile comatose children, increasing the possibility of favorable clinical outcomes.