To create code that expedites the analysis of electroencephalography (EEG) reports by detecting and automatically tabulating epileptiform abnormalities (EAs).

EAs are important in seizure/epilepsy detection and in acute brain injury monitoring like delayed cerebral ischemia after subarachnoid hemorrhages (Kim et al., 2017; Chen et al., 2022, Punia et al., 2015). However, reviewing EEG reports is time-consuming and error-prone; thus, we created an algorithm for rapid EEG report EA annotation.

We manually reviewed 1274 reports for EAs: seizures, sporadic epileptiform discharges, generalized and lateralized periodic discharges (GPDs, LPDs), lateralized rhythmic delta activity (LRDA), bilateral independent periodic discharges (BIPDs), brief potentially ictal rhythmic discharges (BIRDs), and bilateral independent rhythmic delta activity (BIRDA). We also reviewed for generalized rhythmic delta activity (GRDA). We compiled 104 positive and 240 negative phrases indicating EA+GRDA presence/absence, which were then used to build code that phrase matched and tabulated findings in each report. 80% of data was used for training, 20% for testing. Due to conflicts in reports and phrasing ambiguities, a “needs review” category was created for secondary manual review. 39 ambiguous phrases triggered this secondary review (ex: GRDA cannot be excluded).

The algorithm had 1.56% reports with errors identifying EAs+GRDA. 62.11% were accurately matched without need for secondary review. 36.33% were marked as needing review. Of those, 53.76% were marked for potential conflicts stating both the presence and the absence of an EA+GRDA, 26.88% for “periodic” or “rhythmic” term use without localization specified (generalized vs. lateralized), 2.15% for phrases that could indicate multiple forms of EAs+GRDA, requiring manual clarification, and 17.20% for a coding error (later rectified).

These results suggest that this algorithm can automate EAs+GRDA extraction and expedite review of EEG reports. This tool allows researchers to assess the relationship of EAs+GRDA in large datasets more easily in future research.