Artificial Intelligence: Fundamentals and Breakthrough Applications in Epilepsy

Abstract

Artificial intelligence, machine learning, and deep learning are increasingly being used in all medical fields including for epilepsy research and clinical care. Already there have been resultant cutting-edge applications in both the clinical and research arenas of epileptology. Because there is a need to disseminate knowledge about these approaches, how to use them, their advantages, and their potential limitations, the goal of the 2023 Merritt-Putnam Symposium and of this synopsis review of that symposium has been to present the background and state of the art and then to draw conclusions on current and future applications of these approaches through the following: (1) Initially provide an explanation of the fundamental principles of artificial intelligence, machine learning, and deep learning. These are presented in the first section of this review by Dr Wesley Kerr. (2) Provide insights into their cutting-edge applications in screening for medications in neural organoids, in general, and for epilepsy in particular. These are presented by Dr Sandra Acosta. (3) Provide insights into how artificial intelligence approaches can predict clinical response to medication treatments. These are presented by Dr Patrick Kwan. (4) Finally, provide insights into the expanding applications to the detection and analysis of EEG signals in intensive care, epilepsy monitoring unit, and intracranial monitoring situations, as presented below by Dr Gregory Worrell. The expectation is that, in the coming decade and beyond, the increasing use of the above approaches will transform epilepsy research and care and supplement, but not replace, the diligent work of epilepsy clinicians and researchers.

Introduction

Artificial intelligence (AI), machine learning (ML), and deep learning (DL) are having major impacts on research and on clinical care in various medical fields. ^{1 -4} This also applies to the field of epilepsy ^5,6 making it imperative for both clinicians and researchers to acquire additional knowledge about these approaches. In this article, we present the current and expected impact of these approaches on various areas of epileptology.

Fundamentals and Principals of ML and AI

Machine learning and AI will not replace clinicians and scientists, but clinicians and scientists assisted by ML/AI will replace clinicians and scientists without ML/AI. Therefore, clinicians and scientists who diagnose, treat, and research epilepsy should aim to understand both the benefits and key limitations of these technologies.

A ML tool uses data-driven methods to accomplish a specific task by learning the relationship between input data and a variable it seeks to predict (eg, predicting if a patient has epilepsy based on outpatient electroencephalography [EEG]). Deep ML is a subtype where there are multiple layers of processing the input data to predict the variable of interest. With infinite data, deep ML will perform at least as well as single-layer ML, but there are key challenges in obtaining needed large, quality datasets. In contrast, AI uses data-driven methods to be able to perform a broad array of tasks, including tasks for which the algorithm has not been specifically trained. Generative AI can generate content, including but not limited to text and images.

There are numerous metrics to evaluate the performance of ML/AI. While sensitivity and specificity are valuable metrics because they are not influenced by the prevalence of epilepsy in the data, positive and negative predictive value may be more clinically relevant. The clinical relevance of predictive values is shown within Calibration Curves, which highlight the probability of the target outcome based on the prediction of the ML/AI tool. For example, the 2HELPS2B score predicted the likelihood of seizures on further continuous inpatient EEG monitoring up to 72 hours. ^7,8 A 2HELPS2B score of 0 indicated less than 5% chance of seizures in further monitoring. In contrast, a 2HELPS2B score of 2 or higher indicated a 27% chance of future seizures, which would reduce to less than 5% if there were no seizures on at least 24 hours of seizure-free continuous EEG.

For specific applications of ML/AI to seizures, other metrics of performance likely are relevant. When a device is predicting a patient’s real time risk of seizure, the percent of time in high warning is important because there is less practical value of predicting high risk 99% of the time. ⁹ Additionally, ML/AI algorithms also must reliably indicate when the input data was insufficient to make a prediction; for example, when a battery is empty. ¹⁰

In addition to these metrics, we must recognize that ML/AI are data-driven algorithms that are trained on finite datasets. Tools trained based on high quality datasets with input features selected based on experts’ valid predictions may require as few as 20 examples. However, most tools aim to address challenging clinical problems without clear solutions. Therefore, training these algorithms may require collection of hundreds or thousands of possible input features from thousands or millions of patients. A challenge of obtaining data at this scale is the requirement for gold-standard labels. Under the principal of garbage in–garbage out, if we train an ML/AI to identify garbage labels, then the output of the ML/AI also is garbage.

To integrate these powerful tools into clinical practice, clinicians must be alert of other key limitations. Machine learning/AI are trained based on historical data; therefore, its predictions likely will perpetuate historical discriminatory practices. Specifically, even if an ML/AI is not directly given race information as input, it could use other systematic determinants of health to perpetuate the historical bias that black patients were less likely to undergo surgical treatment of medication-resistant epilepsy. ^11,12 Additionally, screening must occur for adversarial examples where the ML/AI’s prediction is wildly incorrect, despite a seemingly insignificant changes in the input data. ^13,14 Further, hallucinations are outputs of an AI where it is making a prediction based on data that it has seen, but that prediction may appear false when viewed by human experts. However, hallucinations may be both a bug and a feature: these predictions may initially appear false but they may represent information not yet appreciated by human experts.

Lastly, ML/AI tools cannot replace human clinicians yet. To harness the power of ML/AI, human experts must be able to interpret the output of these tools in the context of other subjective and objective data. In clinical neurology, humans also must translate this complex data analysis into clinically relevant decisions that should be made in collaboration with the patient. The goal of ML/AI is not to replace the human expert. Instead, ML/AI can allow human experts to make better decisions in less time.

Screening for Medications in Neural Organoids

Brain organoids are 3D tissue cultures, grown from pluripotent stem cells (PSC) offering an in vitro platform to mimic the brain cytoarchitecture, reaching neuronal maturation following the human brain development milestones. ^{15 -18} The main innovation of the model is the possibility to derive brain organoids from patient derived induced PSC (iPSC), reproducing reliably the genetic and genomic features of each individual. ¹⁹ Brain organoids have been used to model the underlying molecular pathogenesis from multiple neurological disorders, as reviewed in detail elsewhere. ¹⁹ Brain organoid rise over other non-human brain models due to their ability to reproduce the specific features of human brain development, such as the expansion of neuroepithelial progenitors and outer radial glia progenitors (oRG) or the neurogenic potential in the human cortical subplate. ^20,21 The protracted neurogenesis in humans not only has impact on the generation of cortical excitatory neurons but also on the origin and expansion of interneurons, which account for 30% of the total number of neurons in the human cerebral cortex, while in mouse they barely reach 15%. ^{22 -24} The use of brain organoids has helped in delineating the underlying molecular mechanisms involved in the acquisition of human-specific traits. ^{25 -28} In parallel, some of these studies centered in the human brain organoid model have shed light in the description of the pathogenesis of multiple epileptic disorders such as Timothy syndrome ²⁸ and Tuberous sclerosis. ²⁹ Yet, modeling epilepsies requires a robust neuronal network, a feature hampered by the intrinsic limitations in the generation of human brain organoids. On the one hand, the in vitro culture conditions depleted of vasculature induce stress in a subpopulation of cells located at the inner core of the organoid. ^30,31 On the other hand, brain organoids are morphologically variable, jeopardizing the assessment of a clear disease-associated phenotype. The organoid variability can be reduced, but not eliminated, using regional guided differentiation protocols. ^15,18,28 However, these models will only shed light in the phenotype generated intrinsically in those regional brain compartments. Opposite, the use of unguided brain organoids differentiation will allow a more extensive phenotype characterization, but the interorganoid variability will be exacerbated. Most of the variability observed in both guided and unguided organoids derives from two main organoid differentiation processes: (a) the heterochronic intraorganoid neurogenesis (multiple patched RG neurogenic rosettes in each organoid) and (b) the neuronal circuitry network that grows and forms depleted of structured area signaling. ^18,32,33

Overcoming these limitations is needed to extend the use of brain organoids in modeling all epilepsies, including traumatic and idiopathic, although it is not an easy task. It requires the incorporation of multiple fields, and certainly the use of sophisticated computational analysis that can process the huge volume of data derived from brain organoids cultures, especially when longitudinal analysis is used. Currently, DL is being used to compute the underlying gene regulatory networks in brain organoids, ³⁴ transcriptomic data, ³⁵ and establishing trajectories in the differentiation process of the multiple subtype analysis. ^36,37 Moreover, efforts are underway to generate DL models to analyze live imaging data of longitudinal studies that can reliably identify the phenotypic outcome of disease associated organoids. ³⁸ The intertwined advances in microscopy, computational/quantitative biology, and bioengineering are promising for moves toward the implementation of organoids and of connected DL methods, not just as a preclinical model but also as a personalized medicine tool.

Prediction of Clinical Response to Medication Treatments

Anti-seizure medications (ASMs) are the mainstay of treatment for epilepsy. Patients have variable response to ASMs in terms of both efficacy and tolerability. For a patient with newly diagnosed epilepsy, a “trial-and-error” process of ASMs prescription is undertaken with the aim to achieve seizure freedom, generally defined as no seizure for at least 12 months. During this process critical time is lost for many patients in trying ineffective ASMs until the “right” one is found. Patients have described this period of waiting to achieve seizure freedom as a time of vulnerability, uncertainty, and confusion with a sense that their lives are “on hold.” ³⁹

Further, recent studies have shown that the proportion of patients with drug-resistant epilepsy (DRE) remains steady, despite the expanded pharmacological options. Patients with DRE should be considered for non-drug options which include resective surgery, neuromodulation, and dietary treatments. However, there is no reliable way to predict whether a patient has DRE at the outset. A more reliable way to predict treatment response and DRE for the individual patient is needed. One potential approach is the use of AI/ML models.

To date, 8 studies that developed ML models to predict ASM response have been published (Table 1). Most studies used clinical information, supplemented by genetic and EEG information in some studies. Most studies have employed traditional ML techniques although 2 recent studies also investigated DL methods and are discussed further. De Jong and colleagues, in 2021, ⁴⁰ combined both clinical and genetic data from patients recruited in two phase 3 clinical trials to develop ML models to predict the response to adjunctive brivaracetam. Among the tested models, the gradient-boosted trees classifier, a traditional ML model, trained jointly on all data modalities, achieved the best performance, with an AUC of 0.76 in the training dataset and 0.75 in the validation dataset. Of note, the Area Under the Curve (AUC) of the model trained on clinical features alone was 0.71.

Table 1.

Studies That Developed Machine Learning Models to Predict Response to ASMs.

First author (year published)	Data source	Number of patients	Predicted outcome	Features	Models tested	Validation
Petrovski (2009), 41 Shazadi (2014) 42	Australia and UK cohorts	1058	ASM response in patients with newly diagnosed epilepsy	Genetics	kNN	Internal + external (not validated)
Devinsky (2016) 43	Claims database	34 990	ASM change	Administrative	RF	Internal (separate)
Ouyang (2018) 44	Single center	20	50% seizure reduction upon ASM change	Quantitative EEG	SVM	Internal
Zhang (2018) 45	Single center	46	Seizure freedom with levetiracetam as first monotherapy	Clinical, EEG	SVM	Internal
Yao (2019) 46	Single center	287	ASM response in patients with newly diagnosed epilepsy	Clinical	XGBoosta, DT, RF, SVM, LR	Internal
De Jong (2021) 40	Two phase 3 trials	235 and 47, respectively	Response to adjunctive brivaracetam	Clinical, genomics	GBTa, sparse multi-block PLS-DA, MNN, elastic net, LR	External
Hakeem (2022) 47	Multicenter	1798	Response to first prescribed ASM	Clinical	Transformera, MLP, LR, SVM, XGBoost, RF	Internal + external

Abbreviations: ASM, anti-seizure medication; DL, deep learning; DT, decision trees; GBT, gradient-boosted trees; kNN, k-nearest neighbors; LR, logistic regression; ML, Machine learning; MLP, multilayered perceptron; MNN, multimodal neural network; PLS-DA, partial least squares discriminant analysis; RF, random forest; SVM, support vector machine.

^a Best performing model; regular font denotes traditional ML models and bold font denotes DL models.

Hakeem and coworkers in 2022 ⁴⁷ developed and validated a deep learning model to predict the response to the first ASM prescribed. A total of 1798 adults with newly diagnosed epilepsy seen at epilepsy centers in Scotland, Malaysia, Australia, and China were included. Sixteen clinical features were used to develop different ML models to predict the probability of treatment success, defined as no seizure while taking the first ASM prescribed during the initial 12 months of treatment. The transformer model, a DL model, outperformed 5 traditional ML models with an AUC of 0.65 and a weighted balanced accuracy of 0.62 on the test set.

Even fewer studies have applied ML to predict DRE (Table 2). Of the 3 studies published, all tested traditional ML methods, using different definitions of DRE, and none were externally validated.

Table 2.

Studies That Developed Machine Learning Models to Predict Drug-Resistant Epilepsy.

Study	Data source	Number of patients	Predicted outcome	Features	Models tested	Validation
Silva-Alves (2017) 48	Single center	237	Drug-resistant mesial TLE	Clinical and genetics	RF	Internal
An (2018) 49	Claims database	292 892	Number of ASMs prescribed	Administrative	SVM, LR, RF	Internal
Delen (2020) 50	Electronic medical records	37 024	ASM vs non-ASM use	Clinical	DT, RF, GBT	Internal

Abbreviations: ASM, anti-seizure medication; DT, decision trees; GBT, gradient-boosted trees; LR, logistic regression; RF, random forest; SVM, support vector machine; TLE, temporal lobe epilepsy.

The application of ML in prediction of ASM response is still in its infancy. However, emerging studies have shown promising results even by including clinical information alone. There is expectation that the incorporation of other data modalities relevant to treatment response, in particular genomics, neuroimaging, and EEG, will enhance the accuracy of the models. It is envisaged that, rather than replacing clinical decision, ML models may be used as clinical decision support tools, for example in selecting ASMs or fast-tracking presurgical evaluation for patients predicted to have DRE. This will hopefully shift the current paradigm from “trial-and-error” to one of personalized medicine, leading to improved health outcomes. In this way, through the use of AI, observations made in clinical practice would become “interventional.”

The Role of AI in EEG Analysis

Epilepsy, a neurological disorder characterized by recurrent seizures, affects millions of individuals worldwide. The primary diagnostic tool for epilepsy is the electroencephalogram (EEG) that records the brain’s electrical activity. The EEG is commonly recorded from the scalp (scalp-EEG) to capture interictal epileptiform spikes, sharp waves, and seizures. For people with focal DRE invasive devices for brain sensing (iEEG) and electrical stimulation are widely used for both diagnostic and therapeutic purposes. Traditional EEG analysis relies on visual inspection by skilled clinicians, a process that is time-consuming and subject to inter-observer variability. Artificial intelligence has opened new horizons in non-invasive scalp-EEG analysis and the iEEG sensing enables seizure tracking and adaptive feedback control with chronically implanted devices. It is widely believed that the field of epileptology is at an inflection point and soon to realize revolutionary advances from AI that will transform the diagnosis and treatment of epilepsy. Here we focus on some recent advancements in ML and AI-enabled EEG in real-world health care settings.

The scalp-EEG is a cornerstone of clinical epileptology. Investigations using quantitative methods applied to scalp-EEG have a long history, including various ML, DL, and feature engineering-based approaches. However, most studies were previously limited by relatively small data sets, limited prospective testing, and with little information on the generalizability of the approach.

A recent study by Tveit et al ⁵¹ on routine outpatient scalp-EEG is notable for its development of an AI model designed for the automated interpretation of routine EEGs. The model was trained on over 30 000 EEG recordings and validated across multiple independent datasets. The AI model demonstrated high diagnostic accuracy on a large hold out dataset from different institutions, achieving human expert-level performance.

Scalp-EEG is also a critical diagnostic tool for evaluation of patients in the intensive care unit. A recent study by Jing et al, ⁵² created a DL AI algorithm that was trained on over 6000 scalp-EEGs from 2711 patients. The model demonstrated expert-level reliability in identifying seizures and seizure-like events and advanced an interesting concept of “ictal-interictal-injury continuum” of patterns, including seizures, periodic discharges, and rhythmic delta activities. The performance in classifying these events matched or surpassed human experts based on calibration and discrimination metrics, indicating its potential as a tool for the expedited review of EEGs. The use of iEEG for mapping physiological and pathological epileptogenic brain using ML has a long history, but to date largely has been limited to relatively small number of patients. ^53,54

Applications to invasive EEG recorded with chronically implanted devices are moving from open-loop systems to sophisticated brain coprocessors that integrate neural sensing and stimulation with cloud computing and ML algorithms. Classification algorithms fine-tuned through extensive ambulatory data for detection of epileptiform spikes and seizures, ⁵⁵ and classification of sleep-wake behavioral state, ⁵⁴ have been demonstrated in ambulatory subjects living in their natural home environments. These devices can be coupled with patient-reporting and behavioral tracking and promise to personalize and provide rapid optimization of neuromodulation therapy.

Conclusions and Overview

In conclusion, one can indicate that in epilepsy AI applications are proving to be of tremendous potential for enhancing research and clinical practice. This includes hypothesis generation, organizing/analyzing literature and data, understanding mechanisms, directing therapy, analyzing neurophysiological and neuroradiological data, and predicting responses to therapy and outcomes of surgery, as discussed above and elsewhere. ^5,6 Additionally, there are currently available, and there will increasingly be available, various increasingly useful AI tools. These include various Chatbot programs, designed to simulate human conversation, such as ChatGPT which utilizes OpenAI’s Generative Pre-trained Transformer (GPT) model, and Dougall GPT which is directed at health care providers for providing well informed and current answers to clinical queries. ^1,2 These and other AI approaches such as large language models, similarity graphs, and knowledge graphs can also aid in knowledge organization, integration, and hypothesis generation, as well as enhancing medical education and writing. ⁴ In addition, AI is proving to be useful in assisting prediction of drug response as presented above and elsewhere. ⁴⁰ Even more importantly, AI techniques will allow integration and analysis of large data sets from consortia and are expected to enhance interpretation of clinical, neuroradiological, and neurophysiological data in research and in clinical decision-making. ^56,57 Importantly, the use of these approaches has to be judicious and must guard against dependence on biased, old or irrelevant data, plagiarism and ethical problems that arise from uncritical or inappropriate use. ³