Abstracts

Rationale:
Artificial intelligence (AI) and machine learning (ML) methods have been increasingly leveraged to analyze brain magnetic resonance imaging (MRI) data to aid with clinical decision-making in epilepsy.¹ Major applications of AI/ML studies include diagnosis, seizure lateralization, lesion localization, and prognosis after surgical intervention. Given the heterogeneity of AI/ML techniques and algorithms, the general accuracy and their effectiveness remains unclear. Here, we aim to assess the performance of current AI/ML models in four major clinical applications: diagnostic ability to identify epilepsy patients from healthy or diseased cohorts, lateralize the epileptic lesion, localize the seizure-onset zone, and predict post-surgical outcomes.




Methods:
In concordance with PRISMA guidelines, a systematic review utilizing MEDLINE and Embase databases was conducted June 2023 yielding 2606 publications (Fig 1A). Studies were included if they used any AI/ML model trained on MRI images to classify patients according to the four stated outcomes and reported any evaluation metric (sensitivity/specificity, AUC under ROC, or total accuracy score). Conference abstracts without a full-text publication or studies with n< 10 patients were excluded. Information on MRI machine, AI/ML algorithm, demographics, and algorithm performance metrics were extracted from all articles. The risk of bias of each study was assessed using the Prediction model Risk of Bias Assessment Tool (PROBAST) tool.²




Results:
200 studies were included with resulting study characteristics shown in Fig 1B. 171 studies reported an accuracy score (diagnosis n= 87/95; localization n=17/35; lateralization n= 37/42; surgical outcome n=20/28) and were, thus, used in the meta-analysis. AI/ML methods in localizing epileptic lesions and surgical outcome prediction showed the lowest accuracy rates of 80% and 84% respectively, while lateralization of temporal lobe epilepsy patients and diagnostic capability showed higher accuracy rates at 90% and 91% (Fig 2B). All estimates had significant heterogeneity between studies. There were no significant differences in using unimodal or multimodal MRI data, suggesting that AI/ML methods are robust across different modalities and algorithm types (Fig 2A). All studies report an overall high risk of bias; all studies on diagnosis, localization, and lateralization had a high risk of predictor bias, while all studies on diagnosis and surgical outcome had a high risk of analysis bias.




Conclusions: Applications of AI/ML in four major clinical applications from diagnosis to prognosis of epilepsy patients show high accuracy rates from 80-91% and are robust across various study designs and algorithm types. However, significant heterogeneity between studies and high risk of bias remains a challenging barrier to clinical translation. Future studies should prospectively image participants prior to their diagnosis or surgery, verify patients’ diagnoses histopathologically, and incorporate model validation using independent, multisite test cohorts.





1. 1 Lucas A, et al. Rev Neurol. 2024 Jun;20(6):319-336.

2. 2 Wolff, Robert F., et al. Annals of internal medicine 170.1 (2019): 51-58.


Funding: NA