At the 75^thAAN 2023 Annual Meeting, Dr. Kate Davis, Director of Penn Epilepsy Centre, University of Pennsylvania, the United States, presented his study on the novel clinical applications of artificial intelligence (AI) in the diagnosis and management of epilepsy.¹ 

Nowadays, AI incorporates several kinds of analysis such as machine learning (ML), which involves neural networks and deep learning (DL).¹ ML can be classified into supervised or unsupervised.¹ In the supervised ML approach, objects of interest are labeled and an algorithm is designed to train the system in identifying objects of interest based on various characteristics, also called the “training phase”.¹ To determine whether the model has been adequately trained to identify objects of interest, the “testing phase” evaluates whether the model can correctly recognize the items of interest when a new set of data is presented.¹ As to the unsupervised learning, the system categorizes objects into different classes based on similar features, in the absence of any labeled data to train it.¹ Artificial neural networks (ANNs) are a subset of ML made up of an input layer that transmits signals into multiple hidden layers, which then further transmits signals into an output layer.¹ Through repeated training, the correct connections between these layers are strengthened, thereby improving the intended identification and classification.¹ If there are multitude of hidden layers in an ANN, it is called DL as it can vastly improve classification and allow for very complex operations.¹

AI has been used to correctly predict seizures 3 days in advance, based on large data sets of patient intracranial electroencephalogram (iEEG) reports in 2021.¹ In clinic settings, a non-invasive device using AI has been adopted to rapidly detect seizures.¹ Personalized treatment plans can be developed by using AI natural language processing tools to scan patients' electronic health records for an array of relatively rare genetic pediatric epilepsy precede diagnoses.¹ This can provide information regarding different forms of seizures and when seizures occur, as well as conduct natural history analysis of various illnesses.¹ It can also classify seizure severity and seizure control as per the types of medication taken, thus aiding in developing a more personalized treatment plan.¹ 

Imaging has been greatly augmented by AI, which can potentially help develop personalized treatment plans.¹ In a study, previously hidden magnetic resonance imaging (MRI)-negative type II focal cortical dysplasia (FCD) has been detected in 85% of the patients using computational neural networks “human-in-the-loop” ML.¹ Imaging has also been used to improve intracranial electrode placement.¹ Resting-state functional MRI and diffusion imaging were used to create a connectome fingerprint of patients who had a good surgical outcome after temporal lobectomy.¹ This fingerprint can be used to compare against new patients and predict their surgical outcomes.¹ 

A large amount of iEEG clip data from many patients can be clustered and rapidly annotated using AI.¹ These data reveal a cluster of patients who respond well to a particular device setting and tailor new patients to get those settings.¹ AI has also been used to identify interictal epileptiform discharges in clinics.¹ Besides, ML has been used to create a normative map of iEEG, which is then compared with the maps of new patients to see which nodes are abnormal and identify the likely seizure foci.¹ 

Post-surgical outcomes can also be predicted by using an AI-based tool that has already been made available in clinics.¹ Using readily available clinical and EEG features in a nomogram, seizure freedom at 2 years can be predicted.¹ Memory performance in temporal lobe epilepsy has also been predicted with structural connectomics using AI as it can detect these signals from a very large data set that the human eyes fail to detect.¹  

Dr. Davis ended by describing the application of AI in randomized controlled trials (RCTs), saying that “AI has a very good potential for improving our approach to RCTs. We can use patients’ history of seizures to predict seizures within the next 24 hours and adopt an AI forecasting approach.”¹ Potentially, fewer patients will be needed and less time will be spent for conducting a successful RCT if this approach is being used. Moreover, trials can be made more patient-specific.¹ Nevertheless, It is crucial to ensure that there are no errors in data collection, and the data provided to AI are free of bias, as these mistakes can be fatal.¹ In fact, currently there are no regulatory guidelines for the use of AI in medicine, and data privacy continues to be a concern as well.¹