Automated Epilepsy Diagnosis using AI-based ResNet152 Deep Learning for Early Detection and Improved Patient Health Outcomes

Abstract

Epilepsy is a long-lasting neurological condition that is characterized by repeated seizures and affects millions of people across the world. Traditional methods of diagnosis that are based on manual interpretation of the clinical examinations and electroencephalogram (EEG) are frequently time-consuming and subjective. Thus, early and correct diagnosis is paramount to successful treatment and better patient outcomes. To improve the efficiency and accuracy of diagnoses, the current study proposes an automated system of epilepsy diagnosis using the ResNet152 deep learning architecture. This was trained and tested on a labelled dataset comprising of epileptic and non-epileptic cases. To enhance the performance of the models, preprocessing methods such as normalization and feature extraction were employed. A classification method based on risk score was used and the model was tested using metrics like accuracy, recall, precision, F1-score, confusion matrix and precision-recall curves with threshold optimization. The findings reveal that there is stable validation accuracy of about 8384, which means that there is a consistent learning behaviour. The confusion matrix indicates that it has a high specificity of 586 correctly identified true negatives but low sensitivity with only 2 true positives and 112 false negatives. Moreover, the precision-recall curve (AUPRC = 0.19) indicates the trade-off between precision and recall as a sign of class imbalance issues. In general, the suggested ResNet152-based framework has a high potential to be used as an automated screening tool, especially to detect non-epileptic cases. Its sensitivity, class balancing and optimization techniques, however, need to be improved to elevate its dependability in the clinical application of detecting early epilepsy.