Authors :
Presenting Author: Majed Alzahrany, MD – King Abdulaziz University
Anas alyazidi, MD – King Abdulaziz University
Ghada Abbas, MD – King Abdulaziz University
Haythum Tayeb, MD – King Abdulaziz University
Abeer Khoja, MD – King Abdulaziz University
Osama Muthaffar, MBBS, SBPN, CSCN, ABCN – King Abdulaziz University Hospital
Ahmed Bamaga, MD – King Abdulaziz University
Fatoon AIshehriy, . – Collage of computer science, King Khalid University
Lama Ayash, . – Collage of computer science, King Khalid University
Renad Alsubaie, . – College of Medicine, King Faisal University, Al-Ahsa
Rationale:
Epilepsy affects over 50 million people worldwide and presents significant diagnostic challenges, making early and accurate identification essential across all age groups. Electroencephalography (EEG) is a cornerstone in epilepsy diagnosis, yet its interpretation is complex and constrained by the global shortage of trained neurophysiologists. Artificial intelligence (AI) has shown promise in medical diagnostics, and its application in EEG interpretation could enhance diagnostic accuracy and reduce delays.
Methods:
This retrospective diagnostic validation study evaluated an AI-based system for automated EEG interpretation in various age group epilepsy. EEGs from 196 patients aged 1-91 years were analyzed, with expert neurophysiologist interpretations serving as the reference standard. The AI model, based on MobileNetV2 architecture, was trained to classify EEGs as normal or abnormal and to identify epileptiform activity into categories relevant for clinical decision-making: epileptiform-focal, epileptiform-generalized, non-epileptiform-focal, and non-epileptiform-diffuse. Performance was assessed using sensitivity, accuracy, area under the ROC curve (AUC), and Cohen’s kappa coefficient for interrater agreement. Clinical metadata were integrated to enhance model performance.
Results:
The AI model achieved a validation accuracy of 54% and sensitivity of 61% when clinical metadata (e.g., age, sex, artifact presence, sleep state) were included and outperforming image-only models (best accuracy: 47%). The model demonstrated moderate discriminative ability (AUC = 0.62; 95% CI: 0.58–0.66) and fair agreement with expert interpretations (κ = 0.35, p < 0.001). Agreement decreased in artifact-heavy EEGs (κ = 0.22). Abnormal EEGs were correctly flagged in majority of cases, including those with focal slowing, generalized slowing, and epileptiform discharges.