Epilepsy is a disease characterized by seizures that The International League Against Epilepsy (ILAE) defines as “a transient occurrence of signs and/or symptoms due to abnormal excessive or synchronous neuronal activity in the brain” [1]. It is estimated that 60% of all epilepsies are focal epilepsies [2] in which the abnormal neural activity starts locally and spreads by engaging other pathological and healthy brain areas [3], [4].

The most prevailing treatment of epilepsy is administering antiepileptic medication to suppress seizure generation and propagation (for a recent review of antiepileptic medication, see [5]). However, almost 30–40% of patients are pharmacoresistant [6], [7]. Moreover, the risk of pharmacoresistance is higher in patients with focal seizures compared to those with generalized seizures [8]. For those patients presenting with focal seizures, curative surgery to remove or inactivate the epileptogenic focus remains the only solution to stop seizures occurring [9], [10].

However, these procedures are associated with substantial clinical and cognitive risks. Post-surgical clinical outcome refers to success in managing seizure status after the surgery. For instance, Mohan et al. [11] reported that 47% of patients remained seizure-free at five and 38% at ten years. However, these curative surgeries have cognitive risks as they can lead to changes in neuropsychological functioning, impacting patients’ quality of life [12], [13]. A meta-analysis performed by Sherman et al. [14] showed that 44% of patients with left temporal lobe epilepsy faced verbal memory disorders, and 34% experienced a naming decline after surgery.

Given multiple risks of epilepsy surgery, each surgical candidate undergoes a comprehensive neuropsychological and neuroimaging assessment before surgery. The main goal of this evaluation is a cost–benefit analysis that reveals the risks of cognitive decline versus potential seizure freedom [12], [15]. Accordingly, there has been a considerable effort to predict and prognosticate clinical and cognitive outcomes of epilepsy surgery using patients’ preoperative characteristics [12], [16], [17], [18]. Indeed, numerous studies aimed to predict the post-surgical outcome in epilepsy patients (herein ESOP – epilepsy surgery outcome prediction) based on patients’ preoperative characteristics such as clinical [19], [20], [21], cognitive [19] and cerebral data provided by anatomical [20], [22], [23], functional neuroimaging [24] and EEG recordings [20], [25].

However, these studies yielded different results. For instance, some studies reported the duration of epilepsy to be predictive of postoperative seizure freedom [26], [27], while others did not [28], [29], [30]. Similarly, some studies in epilepsy patients found preoperative verbal memory to be predictive of postoperative decline of this function [31], [32], while others did not [30]. Furthermore, some researchers found cortical language mapping to predict postoperative naming decline in temporal epilepsy patients [24], but not others [33].

One solution to overcome these differences is performing meta-analyses and systematic literature reviews to systematically synthesize and merge the results of different studies with the same topic to reach robust and valid empirical evidence [34], [35], [36]. Indeed, there are various meta-analyses of studies on ESOP [10], [36], [37], [38], [39] that tried to reconcile diverse findings by combining and weighting the results of different ESOP studies based on the study quality to find the most robust and valid evidence [40]. These studies mostly focused on post-surgical clinical [10], [38], [41], [42] or cognitive outcomes [14], [39], [43] and identifying valid prognostic features clinical or cognitive epilepsy surgery outcomes [30], [37], [38], [40], [44], [45]. Importantly, one common conclusion of these meta-analyses and systematic reviews is that one of the main reasons for different findings across ESOP studies are differences in employed methodology [37], [40], [45]. For instance, a characteristic that differentiates between ESOP studies is the number of participants used to make the prediction model. This is a crucial point given that small sample sizes can lead to sampling bias [46]. Indeed, Armon et al. [29] warned that large sample sizes are needed (preferably from multiple institutions) for reliable identification of predictors for epilepsy surgery and external validation of these predictive models. Moreover, with the development of artificial intelligence, there is an increased interest in applying machine learning (ML) approaches for predictive studies of epilepsy [47]. This trend could be explained by ML's ability to identify complex relationships between large, multimodal data sets and parameters [47], [48], [49]. This is particularly true for epilepsy surgery since physicians (neurologists and neurosurgeons) base surgery decisions on multimodal data obtained with various neuroimaging methods [functional magnetic resonance imaging (fMRI), positron emission tomography (PET), electroencephalography (EEG)] and neuropsychological assessments for every preoperative evaluation. Consistent with this practice, studies show that leveraging the information obtained through different modalities may help to highlight new patterns difficult to observe or effects too weak to be detected, if only one isolated modality is used [50].

Nevertheless, there has been only one systematic literature review that dealt with methodology of ESOP studies. In 2001, McIntosh et al. [36] analyzed methodological approaches of ESOP studies of that time. These authors concluded that ESOP studies face several issues such as small sample size and variability of seizure freedom definition. Recently, Yuan et al. [48] provided a valuable review of ML applications in diagnosis and surgery prognosis for epilepsy patients. However, their review is not systematic, and it focuses only on ML and neuroimaging techniques, not addressing the methodological variability of these studies. Another point that has not been considered by the existing ESOP meta-analyses and systematic reviews is the clinical applicability of the findings. As pointed out recently by Busch et al. [31], although there are a number of models for predicting memory decline after temporal lobe resection, they are not adapted for widespread clinical use and are not easily accessible to physicians.

This narrative descriptive review aims to fill in that gap and provide a systematic and comprehensive synthesis of the methodology used in ESOP studies.

We specifically focused on the methodology used in those studies with the following questions:

what type of data is used in those studies as predictors of epilepsy surgery outcome, and is there a trend to shift from unimodal or multimodal datasets?

what were the characteristics of samples, outcome variables and analytical approaches in the ESOP studies?

how many ESOP studies provide tools for the clinic reinforcing thus a research-practice bridge?

We formulated the following hypotheses related to each goal:

with the development of neuroimaging techniques and their increased application in clinics [51] and neuroimaging studies [52], we expected that the multimodality of predictors in ESOP studies are increasingly reported;

given the recommendations [36] and the development of power calculation tools (e.g., [53], [54]), we expected the sample size of ESOP studies is growing with time. We anticipated that the ESOP studies focusing on clinical outcomes use standardized categories, following recommendations [36], [55], [56] and widely excepted taxonomies [1], [57], while those focusing on cognitive outcomes use metrics that avoid measurement error [55]. Moreover, in line with our expectations regarding larger and richer datasets [49] and the potential of the ML approaches for data integration [47], [48], we expected this approach to be more dominant in newer and multimodal ESOP studies;

we expected to identify a significant number of ESOP studies that provide clinical tools, as they are motivated by a practical question and goal (predicting the outcome of epilepsy surgery in day-to-day practice).

The present literature review is useful for several reasons: (i) it can inform future predictive studies about the available methods and models; (ii) it suggests that the change of methodology over the years should illustrate a shift in the conceptualization of epilepsy (for instance the increase of network measures related to modern connectomic perspectives [23], [58]); and (iii) it informs physicians who have difficulties identifying potential epilepsy surgical candidates [59] on the studies that provide practical tools which (iv) could hopefully encourage future studies towards limiting the gap between cognitive neuroscience evidence and its application in clinical settings. A bridge between these two fields of knowledge can contribute to more adapted individual care and better surgical outcomes. Following this notion, Box 1 presents reminders and precision of some methodological and technical terms used in the present paper that may only be familiar to some readers.