Artificial intelligence‐assisted categorization of interictal abnormalities in adult long‐term electroencephalographic monitoring: Comparable diagnostic accuracy with marked efficiency gain

Abstract

Abstract Objective Artificial intelligence (AI) classifies routine electroencephalograms (EEGs) with expert‐level accuracy, but interpretation of long‐term EEG monitoring (LTM) remains reliant on time‐intensive visual review. We evaluated whether autoSCORE, a clinically approved implementation of SCORE‐AI integrated into an EEG reader, achieves expert‐level diagnostic accuracy for LTM interpretation and whether expert‐supervised use reduces review time versus conventional analysis. Methods A fixed‐and‐frozen AI model was tested on an independent single‐center dataset of 78 adult patient LTM recordings (52 normal, 26 abnormal), with epilepsy monitoring unit discharge diagnosis as the external reference standard. The AI model was benchmarked against two expert panels (three readers each); Panel 1 reviewed autoSCORE EEG detections (hybrid approach), and Panel 2 interpreted full recordings (conventional approach). Diagnostic accuracy and time to diagnostic decision were assessed for fully automated, hybrid, and conventional approaches. Interrater reliability (Gwet's AC1) was evaluated for hybrid and conventional interpretation. Results Interrater reliability was significantly higher for hybrid than conventional approach (AC1 = .81, 95% confidence interval [CI] = .73–.89 vs. AC1 = .43, 95% CI = .32–.53; p < .001). Review time was markedly reduced (median [interquartile range (IQR)] = 2.2 [.5–4.3] vs. 52.4 [33.2–62.9] min, mean [95% CI] = 2.9 [2.5–3.2] vs. 50.8 [46.5–55.2] min, p < .001), a 95.8% median reduction (IQR = 95.6–96.3). Diagnostic performance comparisons were adjusted for multiple comparisons using the Holm–Bonferroni procedure across all 63 tests. The hybrid approach showed significantly higher sensitivity than conventional interpretation (.88, 95% CI = .77–.97 vs. .69, 95% CI = .57–.81; adjusted p = .006). Its numerically higher diagnostic accuracy versus conventional interpretation (.75, 95% CI = .66–.83 vs. .68, 95% CI = .61–.74) was nonsignificant after correction (adjusted p > .99). Hybrid approach diagnostic accuracy was significantly greater than fully automated analysis (.75, 95% CI = .66–.83 vs. .55, 95% CI = .44–.65; adjusted p < .001). Significance Expert oversight of AI‐assisted categorization of interictal epileptiform and nonepileptiform EEG abnormalities improves interrater reliability and efficiency, increases diagnostic sensitivity versus conventional interpretation, and increases diagnostic accuracy versus fully automated analysis in adult LTM.