Can we predict surgical outcomes: A systematic review and critical appraisal of clinical prediction models in epilepsy surgery Authors: Alyssa A. Federico, Mandavi Kashyap, Chantelle Q. Y. Lin, Karl M. Klein, Olayinka I. Arimoro, Samuel Wiebe Published: 2026 ABSTRACT AbstractObjectivePrediction models are increasingly being sought in epilepsy surgery to predict postoperative outcomes and support clinical decision‐making. Studies summarizing the evidence in this area can provide insight into the type of surgical prediction models, their methodology, and their performance and inform areas for future research. Our aim was to address these knowledge gaps through a comprehensive systematic review of prediction models in epilepsy surgery.MethodsA systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses guidelines using four databases. Papers were included if they were primary research studies, human‐based, studied adult or pediatric populations, studied people with epilepsy undergoing surgical management, and developed or validated a multivariable tool to predict epilepsy surgery outcomes. Data extraction was reviewed in triplicate, and the quality of evidence in each paper was assessed using the Prediction Model Risk of Bias Assessment Tool.ResultsThe literature search yielded a total of 11 614 papers, with 42 papers and 113 prediction models included in the final analysis. The median area under the curve and accuracy for all models were .75 (interquartile range = .68–.83) and .76 (interquartile range = .69–.83), respectively. Overall, 54.0% of models underwent internal validation, and 20.4% underwent external validation. Models of cognitive–language outcomes seemed to perform better than those for other outcomes. Overall risk of bias was high in 81% of models, with weakest performance in outcomes and analyses, but trended toward improvement over time. Concerns for applicability were low in 89% of the models.SignificancePrediction models in epilepsy surgery are rapidly proliferating, but most lack external validation, and many still exhibit a high risk of bias. Therefore, caution is needed when interpreting and applying these predictive tools. Evidence of improvement in methodological quality holds promise for enhancing patient care, if coupled with improved model performance. Key points A total of 113 prediction models were identified in 42 papers, primarily assessing seizure outcomes and cognitive and language deficits.Model performance varied across models, outcomes, and methods, and few studies produced clinically applicable prediction tools.Most studies had an overall high risk of bias, with some evidence of improvement over time.Clinical prediction models have the potential to enhance patient care, but rigorous evaluation is critical prior to clinical application. Key points A total of 113 prediction models were identified in 42 papers, primarily assessing seizure outcomes and cognitive and language deficits.Model performance varied across models, outcomes, and methods, and few studies produced clinically applicable prediction tools.Most studies had an overall high risk of bias, with some evidence of improvement over time.Clinical prediction models have the potential to enhance patient care, but rigorous evaluation is critical prior to clinical application. A total of 113 prediction models were identified in 42 papers, primarily assessing seizure outcomes and cognitive and language deficits. A total of 113 prediction models were identified in 42 papers, primarily assessing seizure outcomes and cognitive and language deficits. Model performance varied across models, outcomes, and methods, and few studies produced clinically applicable prediction tools. Model performance varied across models, outcomes, and methods, and few studies produced clinically applicable prediction tools. Most studies had an overall high risk of bias, with some evidence of improvement over time. Most studies had an overall high risk of bias, with some evidence of improvement over time. Clinical prediction models have the potential to enhance patient care, but rigorous evaluation is critical prior to clinical application. Clinical prediction models have the potential to enhance patient care, but rigorous evaluation is critical prior to clinical application. INTRODUCTION Clinical prediction models have proliferated in epilepsy research in recent years, driven by expanding access to data and advanced analytical tools. These models hold the promise of enhancing clinical care by supporting diagnostic and prognostic decision‐making, while also informing actions taken by health researchers, guideline developers, and policy makers. 1 , 2 In the surgical setting, prediction models aim to generate reproducible, precise estimates of specific surgical outcomes for individual patients by applying machine learning or traditional statistical methods to a wide range of predictors, thus simplifying clinical prediction. 1 , 3 Recognition of clinical prediction models as a specific type of research requiring specific reporting and methodological standards led to development of the TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) statement, published simultaneously in 11 high‐impact journals in 2015 to improve the rigor of model development and reporting. 4 Similarly, the Prediction Model Risk of Bias Assessment Tool (PROBAST) was developed in 2019 specifically to assess risk of bias and applicability of primary prediction model studies, an essential aspect of prediction models. Estimating the probabilities of surgical risks and favorable outcomes plays a crucial role in selecting candidates for epilepsy surgery. These are typically discussed by multidisciplinary teams and estimated based on parameters such as pathogenesis, seizure frequency, location of the epileptogenic zone, and type of surgery. 5 Multidisciplinary discussions are the lynchpin of epilepsy surgery decision‐making but often yield subjective assessments rather than objective or probabilistic outcome estimates. Prediction models have the potential to bridge this gap by providing reliable estimates of important clinical outcomes for individual patients, such as seizure recurrence, antiseizure medication (ASM) withdrawal, complications, and neuropsychological deficits. 6 , 7 , 8 A recent systematic review summarized machine learning‐based prediction models in epilepsy surgery, but excluded those based on other commonly used methods, such as statistical models, scales, and nomograms. 9 A comprehensive review of the literature is necessary to assess the state of the art in prediction models in epilepsy surgery, and to inform future efforts to validate and implement these models in clinical practice. Therefore, the overall aim of this study was to perform a systematic review of existing prediction models in epilepsy surgery, assessing risk of bias and applicability and summarizing model performance by type of surgery, surgical outcome of interest, type of model, and model validation. MATERIALS AND METHODS Search strategy This systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) guidelines. 10 The methodology was determined a priori and registered on PROSPERO (Prospective Register of Systematic Reviews; registration number: CRD42022369678). A librarian at the University of Calgary Health Sciences Centre assisted with database selection and generation of the search strategies. Our literature search strategy was designed to be as inclusive as possible to ensure no studies were missed. Embase, MEDLINE, CINAHL (Cumulative Index to Nursing and Allied Health Literature), and Cochrane Library databases were searched using the codes reported in Appendix A. There were no date or language restrictions on the searches, which were limited to human studies only. The original database searches were performed on June 14, 2022, and an updated search was completed on February 22, 2025. Study eligibility The inclusion criteria for the review were as follows: (1) primary research studies, (2) human‐based studies, (3) adult or pediatric population, (4) people with epilepsy undergoing surgical management, and (5) development or validation of a multivariable tool to predict epilepsy surgery outcomes. Surgical management could be in the form of resective or disconnective surgery, minimally invasive surgery (e.g., laser surgery, electrocoagulation), or neuromodulation (e.g., vagus nerve stimulation, intracranial stimulation). Studies were excluded if they met one or more of the following criteria: (1) non‐English papers, (2) non‐primary research studies (e.g. conference abstracts, review papers, response letters, editorials), (3) case reports and case series with <30 patients, (4) animal studies, (5) non‐epilepsy population, and (6) prediction model based on a single variable. Screening All titles and abstracts were screened by two independent reviewers. Articles thought to be relevant or potentially relevant to the review at hand were included in the full text review phase. During full text review, all papers were examined by two independent reviewers to determine their eligibility criteria. If consensus could not be reached, a third reviewer made the final decision. The reference lists of the included papers and related systematic reviews were screened for additional studies. Data extraction All standardized data extraction was performed by two independent reviewers and checked by a third reviewer. The data extracted included information on the study author, year, country, design, purpose, type of study setting, patient population, sample size, number of males and females, number with generalized epilepsy, number with focal epilepsy, number with unknown epilepsy type, and surgery type. The specific information extracted about individual prediction models included the type of prediction model, the type of analysis, the method of choosing predictors, the number of predictors initially considered, the number of predictors included in the final model, the outcome analysis method, the validation method, the measure of association, whether a clinical prediction tool was developed, and the type of tool. For model discrimination, area under the curve, F1 scores, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value were recorded. For model calibration, observed/expected ratios, calibration curves, calibration slope and intercept, and Hosmer–Lemeshow test and Brier test scores were recorded. Clinical outcomes Models were divided into three categories based on the primary clinical outcome of interest: seizure outcome (e.g., seizure frequency or seizure freedom), cognitive and language outcomes (e.g., naming, memory, intelligence), and other outcomes (e.g., ASM withdrawal, postoperative complications, psychiatric comorbidities, patient‐reported outcome measures [PROMs]). Quality assessment The quality of evidence for each model was assessed using PROBAST. 11 The risk of bias (ROB) of each model was assessed across four domains: participants, predictors, outcomes, and analysis. A series of 20 questions related to these categories was rated as having a high, low, or unclear ROB. As recommended by the PROBAST method, the overall ROB was considered low if all question responses were low. If one or more questions were rated high, then the overall ROB was considered high. In studies with one or more responses classified as unclear and no high ratings, the overall ROB was considered unclear. Applicability of the model was assessed separately using three questions and the same rating criteria. Statistical analysis Descriptive statistics were used to denote frequency, proportion, and mean or median values of study characteristics. A meta‐analysis of model performance was not undertaken because of substantial between‐study heterogeneity and incomplete reporting of performance metrics and variances. Instead, area under the curve (AUC) and accuracy, the two most commonly reported model performance metrics, were summarized descriptively as medians with interquartile ranges (IQRs) across all models and prespecified subgroups, and presented as box‐and‐whisker plots, in line with current guidance. 12 As several correlated estimates arose from the same study, medians and IQRs are presented purely as descriptive summaries of the observed distribution of performance metrics and should not be interpreted as pooled effect estimates. To explore the association between reported AUC and other variables, we fitted linear regression models with AUC as the outcome, using study‐level cluster‐robust standard errors to account for clustered data, and performed regression diagnostics through residual distributions and descriptive residual statistics. The percentage of studies with high, unclear, and low ROB in each category according to the PROBAST tool was calculated. To assess whether the scores in each PROBAST domain changed over time, we produced locally estimated scatterplot smoothing (LOESS) curves using kernel‐weighted local polynomial with 95% confidence bands fitted to the ordinal risk of bias scores by publication year. The linear trend was tested by means of linear regression. RESULTS Search results In keeping with our highly sensitive search strategy, the literature search yielded a total of 11 614 papers, with 9190 remaining after duplicate removal, and 8517 excluded based on title and abstract screening, leaving 673 papers for full text review, of which 13 papers could not be obtained. On full text review, most papers were excluded because they were unrelated to epilepsy surgery (n = 354). Other exclusion reasons included absence of a prediction model, wrong population, cohort size < 30, wrong outcome, wrong study design, conference abstract, paper commentary, review paper, or non‐English text. The literature search identified 26 studies, whereas 16 were found through review of reference lists and related systematic reviews. The final analysis included 42 papers reporting 113 prediction models (Figure 1, Appendix B). PRISMA (Preferred Reporting Items for Systematic Reviews and Meta‐Analyses) flowchart of database search and literature review process. CINAHL, Cumulative Index to Nursing and Allied Health Literature. Study characteristics All studies were published between 1998 and 2024, with the highest frequency of publications in the past 5 years (Figure 2A). There were 30 single‐center studies and 12 multicenter studies, including centers in North America (57.1%), Europe (31.0%), Asia–Oceania (26.2%), Latin America (9.5%), and Eastern Mediterranean (2.4%). Most study designs were retrospective cohorts (83.3%), with a few prospective cohort studies (16.7%) and no randomized trials (Table 1). Frequency of published articles by year (A) and number of models published in each article (B). TABLE 1Descriptive statistics of predictive models in epilepsy surgery.Model characteristicsTotal models, n (%)113 (100)Sample size, mean (SD)Per model310 (405)Age, n (%) a Adult46 (40.7)Both25 (22.1)Pediatric38 (33.6)Type of surgery, n (%)Resect–disconnect97 (85.8)Neuromodulation16 (14.2)Predicted outcome, n (%)ASM withdrawal10 (8.8)Surgical complications2 (1.8)Seizure outcome74 (65.5)Cognitive and language25 (22.1)Psychiatric1 (.9)PROMs1 (.9)Type of predictors, n (%)Connectome8 (7.1)EEG, scalp or intracranial8 (7)Heart rate4 (3.5)Imaging, structural or functional18 (15.9)Multimodal72 (63.7)PROM1 (.9)Psychiatric2 (1.8)Number of predictorsPer model, mean (SD)8.6 (9.1)Model output, n (%)Regression formula9 (22.5)Graph/algorithm26 (65.0)Points scale5 (12.5)Validation type, n (%)None29 (25.7)Internal61 (54.0)External23 (20.4)Performance metric, n (%)AUC88 (77.9)Accuracy48 (43.5)ROB predictors, n (%)High19 (16.8)Low38 (33.6)Unclear56 (49.6)ROB outcomes, n (%)High34 (30.1)Low14 (12.4)Unclear65 (57.5)ROB analysis, n (%)High87 (77.0)Low3 (2.7)Unclear23 (20.4)ROB participants, n (%)High6 (5.3)Low95 (84.1)Unclear12 (10.6)ROB overall, n (%)High92 (81.4)Low2 (1.8)Unclear19 (16.8)Concerns applicability, n (%)High8 (7.1)Low101 (89.4)Unclear4 (3.5)Abbreviations: ASM, antiseizure medication; AUC, area under the curve; EEG, electroencephalography; PROM, patient reported outcome measure; ROB, risk of bias. a Not described in two studies. Descriptive statistics of predictive models in epilepsy surgery. Abbreviations: ASM, antiseizure medication; AUC, area under the curve; EEG, electroencephalography; PROM, patient reported outcome measure; ROB, risk of bias. Not described in two studies. Model characteristics Model characteristics are summarized in Table 1 and described in detail for each model in Appendix C. Most papers reported on one or two prediction models, for a total of 113 models included in the final analysis (Figure 2B). The average number of participants used for model development and testing was 310 ± 405, with similar mean values for female (151.8 ± 177.9) and male (164.9 ± 238.5) participants. Most models involved adult participants only (40.7%), 33.6% pediatric participants only, and 22.1% both adult and pediatric participants. Most models focused on resective or disconnective procedures (85.8%), whereas neuromodulation comprised the minority 14.2%. The surgical targets in the 113 included models were temporal lobe (39, 34.5%), frontal lobe (3, 2.7%), mixed temporal and extratemporal (53, 46.9%), multilobar/hemispheric (2, 1.8%), and neuromodulation (16, 14.2%). The most common predicted outcome was seizures (65.5%), followed by cognitive and language outcomes (22.1%). Most models were derived using mathematical or statistical methods (39.6%); fewer were based on scales (26.1%) and machine learning models (24.3%). The average number of predictors included in a final model was 8.6 ± 9.1. Multimodal predictors (clinical, electroencephalography [EEG], imaging) predominated, accounting for 63.7% of all models, followed by magnetic resonance imaging (MRI) features (8.8%), connectome (7.1%), scalp electroencephalography (EEG) data (3.5%), heart rate data (3.5%), other structural imaging features (3.5%), stereo‐EEG data (3.5%), functional MRI data (1.8%), psychiatric variables (1.8%), and PROMs (.9%). Model performance AUC and accuracy were the most frequently reported metrics of model performance and were available in 96.5% of models, of which 77.9% reported AUC and 42.5% reported accuracy. Variances for either metric were infrequently reported. The median AUC for all models was .75 (IQR = .68–.83). When evaluating performance by clinical outcome, the AUC scores were highest for models predicting cognitive and language outcomes (median = .81, IQR = .76–.84), followed by seizure outcomes (median = .74, IQR = .67–.83), and lowest for other outcomes (median = .69, IQR = .670–.730). Performance based on type of validation showed models with no validation had the highest median AUC (median = .83, IQR = .75–.86), followed by those with internal validation (median = .73, IQR = .70–.79); those with external validation tools had the lowest (median = .71, IQR = .61–.81; Figure 3). Models assessing resection–disconnection surgery had higher AUC (median = .75, IQR = .68–.83) than those assessing neuromodulation (median = .73, IQR = .62–.93). There was no association between publication year and AUC values. Distribution of area under the curve (AUC) and accuracy scores by clinical outcome and type of model validation. Boxplots denote median values (red lines), interquartile range (boxes), and range (whiskers). The median accuracy for all models was .76 (IQR = .69–.83). Accuracy was not reported for models predicting ASM withdrawal, postoperative complications, patient‐reported outcomes, psychiatric comorbidities, or scale‐based prediction tools. Median accuracy was higher for cognitive and language outcomes (median = .805, IQR = .730–.850) compared to seizure outcomes (median = .740, IQR = .684–.825; Figure 3). Accuracy by type of model validation showed highest accuracy for those with no validation (median = .81, IQR = .79–.87) and with external validation (median = .83, IQR = .72–.83), and lowest for internal validation (median = .73, IQR = .68–.80; Figure 3). Regarding type of surgery, resection–disconnection had higher accuracy (median = .79, IQR = .70–.82) than neuromodulation (median = .75, IQR = .68–.86). Model validation No validation was undertaken for 25.7% of models, whereas 54.0% of models underwent internal validation and 20.4% external validation. The internal validation methods included resampling techniques (23.0%), split sample (26.2%), and cross‐validation (50.8%). Prediction tools There were 40 clinically applicable prediction models produced from 16 studies (35.4%). These models included algorithmic or graphical tools (23%), regression formulas (8%), and scale‐based prediction tools (4%). Only six models reporting accuracy produced a prediction tool applicable to clinical practice. Risk of bias Most studies had an overall high ROB, despite being applicable to the review at hand (Figure 4). The overall ROB was high in 81.4% of models, unclear in 16.8% of models, and low in 1.8% of models. In contrast, overall concern for applicability was high in only 7.1% of models, unclear in 3.5% of models, and low in 89.4% of models. The least frequently encountered ROB related to selection of study participants, whereas ROB related to analysis methods was the most frequently found. LOESS‐type curves assessing ROB over time demonstrated a trend toward lower ROB over time for all ROB domains, which was statistically significant for ROB participants, ROB predictors, and ROB outcomes (Figure 5). On the other hand, there was no evidence of change over time for the domain concerning applicability, which has been consistently low over time (Figure 5). The detailed PROBAST scores assigned to each paper can be found in Appendix D. Prediction Model Risk of Bias Assessment Tool risk of bias and applicability ratings for all models. ROB, risk of bias. Prevalence of Prediction Model Risk of Bias Assessment Tool (PROBAST) domain by publication year. The panels show the kernel‐weighted local polynomial (locally estimated scatterplot smoothing) with 95% confidence bands fitted to the ordinal risk of bias scores. The central lines represent the estimated mean, and shaded bands represent the 95% confidence intervals. The p‐value is a test of the linear trend in risk of bias (ROB) over publication year from a linear regression model. Outliers Using the distribution of AUC values, we categorized as top outliers those with AUC > .90 (above the 90th percentile), and as bottom outliers those with AUC < .65 (below the 15% percentile). Six models were top outliers, of which four pertained to resective or disconnective surgery and two to neuromodulation. Three used multimodal predictors, two used connectomics, and one used EEG features. Four of these models predicted seizure outcomes, and two predicted cognitive and language outcomes. Notably, only one of the six models underwent external validation. With PROBAST scoring, five of the six models had a high overall ROB and two had high applicability concerns. Eleven models were bottom outliers, of which nine pertained to resective or disconnective surgery and two to neuromodulation. Ten models predicted seizure outcomes and one predicted ASM withdrawal. Ten of the models were derived from multimodal predictors and one from connectomics features. Six of the 10 models underwent external validation, nine had a high overall ROB, and one had high applicability concerns. DISCUSSION Our analysis identified 42 studies reporting on 113 prediction models for various epilepsy surgery outcomes, and 40 clinically applicable prediction tools from 16 studies. No prediction models were identified prior to 1998, and their production has markedly increased in the past 5 years. This explosive production of prediction models in recent years is apparent across all health areas. 13 We found that most models focus on seizure and cognitive–language outcomes, and most use multimodal predictors. There is a dearth of models predicting complications, or PROMs, and no reports addressed minimally invasive surgery outcomes or included patient participation in determining priorities and focus of prediction model development. A key conceptual issue—particularly relevant in epilepsy surgery—is the target population and the decision point at which prediction models are developed and intended to be used. Across the included studies, models are postselection prognostic tools, which predict outcomes among patients already selected for surgery, typically after multidisciplinary/expert consensus. As a result, these models do not inform the upstream question of surgical candidacy (i.e., “Should this patient be offered surgery?”); instead, they estimate the probability of outcomes conditional on having already been selected for surgery. This distinction is crucial when interpreting and applying epilepsy surgery prediction models, because current tools largely support counseling and expectation‐setting after surgical selection rather than decision‐making about whether to recommend surgery in the first place. Although our descriptive analyses of model performance were limited by data availability and should be interpreted cautiously, some insights from the data can inform future exploration. For example, the mean overall AUC (.757 ± .099) fell within the conventional level of acceptable or fair discriminative ability, and values of accuracy metrics approximated those of AUC. 14 Models predicting cognitive or language function had consistently higher performance metrics than those focusing on other outcomes. The superior predictive performance for cognitive outcomes is multifactorial and may relate to better feature selection, as well as more constrained cerebral structures (temporal lobe) and results pertaining to cognitive outcomes, as compared to more complex anatomical structures, networks, and surgical procedures involving seizure outcomes. For example, removing the dominant hippocampus in a patient with a normal MRI will most likely result in cognitive deficits, whereas it will not surely make the patient seizure‐free. Lower predictive ability for models of neuromodulation, compared to resective/disconnective surgery, may reflect a higher degree of complexity and heterogeneity of patients undergoing neuromodulation. Model performance was not consistently associated with use of machine learning versus traditional statistical models; machine learning appeared superior in AUC but inferior in accuracy metrics. This confirms the notion that the best performing machine learning models in one setting may turn out to have the poorest performance on independent validation studies, and that additional artificial intelligence methods seem to contribute little to prediction performance at this time. 15 , 16 Moreover, there is evidence that increasing sample size beyond approximately 1000 seems to have diminishing returns on model performance. 16 Instead, it is proposed that elements such as data quality, data standardization, and improved feature selection may be stronger drivers of performance. 16 Finally, performance metrics tended to be lower for externally validated models, especially compared to those without any validation, emphasizing the crucial importance of external validation. The high proportion of models with no validation (25.7%) or only internal validation (54.0%) is a major concern and should be addressed in future prediction model work. Scientific journals could exert a stronger influence by requiring clear designation of nonvalidated models as tentative or lower grade evidence and encouraging appropriate model validation for designation as more definitive or higher grade evidence. It is worth considering the standard to determine what constitutes acceptable accuracy or AUC thresholds. Acceptable discrimination thresholds (e.g., AUC cutoffs) should be justified against clinically meaningful comparators rather than treated as universal targets, because what constitutes “good” AUC is context‐specific. In epilepsy surgery, one relevant benchmark is unaided clinician judgment. In a study comparing prediction models with clinicians' outcome estimates, there was no evidence that the models outperformed clinicians, underscoring the need to examine whether a model adds value beyond routine practice rather than simply exceeding an arbitrary AUC threshold. 17 On the other hand, other methods of assessing usefulness of predictive models, beyond measures of accuracy or AUC, include their application as aids in clinical decision‐making. One example is the use of decision curve analysis, which can demonstrate predictive benefit across different threshold probabilities, especially if the model shows strong calibration and the range of threshold probabilities is narrower. This application of predictive models in clinical practice requires further study. The high overall ROB in most studies (81.4%) is particularly concerning. The highest ROB was seen in the analysis domain, which involves a lack of adequate calibration, addressing data complexity and avoiding model overfitting. The data support caution in using prediction models for epilepsy surgery outcomes, given their limited predictive ability and risk of bias, and importantly, their unproven generalizability across clinical settings and contexts. However, this finding is not unique to epilepsy surgery prediction models but appears to be pervasive in the field. A meta‐review of 50 systematic reviews that used PROBAST in 1510 different studies on 2104 prediction models also found that 69% of studies had the highest ROB in the analysis domain, and this did not seem to improve over time. 13 On the other hand, it is reassuring that, in our analysis, domains related to participants and predictors had a low ROB overall, and concerns about applicability of the model were low across studies. Moreover, it is encouraging that ROB scores have shown a significant downward trend over time across multiple PROBAST domains, reflecting improved methodological rigor. A main strength of the study is the strict adherence to PRISMA guidelines and triplicate review of all extracted data. 10 These methodological steps ensured scientific rigor and accuracy of the reported data, and our analyses used summary methods recommended for the type of data at hand. Also, our analyses encompassed several subgroups that are relevant to clinicians and researchers. A limitation of the study is the inherently heterogeneous data, with not all studies reporting the same model performance metrics, preventing formal meta‐analyses of the results. Additionally, few studies reported models to predict surgical complications and psychiatric outcomes, decreasing the strength of the results in these two subgroups. Finally, presurgical evaluation has changed over the past decades, and this could have implications for differential predictive ability as better diagnostic tools are used in more recent patient populations. This requires further evaluation in future studies. CONCLUSIONS Prediction models in epilepsy surgery are now widely reported, yet their overall performance is modest, most are judged to be at high risk of bias, and many lack external validation. Accordingly, clinicians making prognoses about individual patients should interpret current model outputs cautiously. Existing models frequently target similar outcomes with overlapping predictors and tend to yield AUC values of approximately .70, so further models of this type are unlikely, by themselves, to advance prognostication. In our review, models predicting cognitive outcomes appeared to perform better and had lower risk of bias than models for other endpoints, suggesting that design features used in this subgroup may warrant closer evaluation and, where appropriate, adaptation to other outcomes. Rather than isolated development of additional models, coordinated initiatives that emphasize common data standards, transparent and improved feature selection, and systematic external validation across centers are more likely to improve clinical usefulness. Areas of emphasis for future studies include rigorous external validation of promising tools, extension to other outcomes such as postoperative complications, PROMs, and postsurgical ASM use, and greater patient involvement in model development. AUTHOR CONTRIBUTIONS Alyssa A. Federico: Study conceptualization; methodology design; investigation; data curation; formal analysis; visualization; writing the original manuscript draft. Mandavi Kashyap: Investigation; data curation; formal analysis; visualization; project administration; review of manuscript. Chantelle Q. Y. Lin: Study conceptualization; methodology design; investigation; review of manuscript. Karl M. Klein: Study conceptualization; methodology design; investigation; review of manuscript. Olayinka I. Arimoro: Study conceptualization; methodology design; investigation; review of manuscript. Samuel Wiebe: Study conceptualization; methodology design; investigation; data curation; formal analysis; visualization; project administration; supervision; writing the original manuscript draft. CONFLICT OF INTEREST STATEMENT None of the authors has any conflict of interest to disclose. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines. Supporting information APPENDIX A Database search strategies. APPENDIX B References for studies included in the final analysis. APPENDIX C Summary of models included in the systematic review. APPENDIX D Detailed Summary of Prediction Model Risk of Bias Assessment Tool scoring for all included papers.