Estimating the likelihood of epilepsy from clinically non-contributory EEG using computational analysis: A retrospective, multi-site case-control study

Abstract

SummaryBackgroundA retrospective, multi-site case control study was carried out to validate a set of candidate biomarkers of seizure susceptibility. The objective was to determine the robustness of these biomarkers derived from routinely collected EEG within a large cohort (both epilepsy and common alternative conditions which may present with a possible seizure, such as NEAD).MethodsThe database consisted of 814 EEG recordings from 648 subjects, collected from 8 NHS sites across the UK. Clinically non-contributory EEG recordings were identified by an experienced clinical scientist (N = 281; 152 alternative conditions, 129 epilepsy). Eight computational markers (spectral [N = 2], network-based [N = 4] and model-based [N = 2]) were calculated within each recording. Ensemble-based classifiers were developed using a two-tier cross-validation approach. We used standard regression methods in order to identify whether potential confounding variables (e.g. age, gender, treatment-status, comorbidity) impacted model performance.FindingsWe found levels of balanced accuracy of 68% across the cohort with clinically non-contributory normal EEGs (sensitivity: 61%, specificity: 75%, positive predictive value: 55%, negative predictive value: 79%, diagnostic odds ratio: 4.64). Group-level analysis found no evidence suggesting any of the potential confounding variables significantly impacted the overall performance.InterpretationThese results provide evidence that the set of biomarkers could provide additional value to clinical decision-making, providing the foundation for a decision support tool that could reduce diagnostic delay and misdiagnosis rates. Future work should therefore assess the change in diagnostic yield and time to diagnosis when utilising these biomarkers in carefully designed prospective studies.Research in ContextEvidence before this studyWe searched Google Scholar and Pubmed (March 21, 2022) for the following phrases ((“EEG” OR “electroencephalogram” OR “electroencephalography”) AND (“biomarker”) AND (“epilepsy” OR “seizure”) AND (“resting state” OR “resting-state”) OR (“normal”)). Several of the existing studies developed deep learning approaches for identifying the presence of interictal epileptiform discharges (IED), with the overarching aim to develop an automated stand-alone diagnostic tool. These approaches are particularly sensitive to the potential presence of artefacts in the EEG recordings and typically include spectral rather than network- or model-based features. We found no studies of more than 100 participants that assessed the cross-validated performance of candidate biomarkers on routine EEG recordings that were clinically non-contributory. One study found near-chance performance of a deep-learning based method using spectral features on a smaller cohort of people suspected of epilepsy (N=33 epilepsy; N=30 alternative conditions) with clinically non-contributory EEGs. Another study found overall accuracy of 69% (N=74 epilepsy; N=74 alternative conditions) but this framework did not use any independent cross-validation methods. Estimates of sensitivity of clinical markers of seizure susceptibility in routine EEG recordings vary between 17-56%. To the best of our knowledge no studies have assessed whether computational biomarkers offer sufficient discrimination between people with epilepsy and an alternative diagnosis to provide potential decision support for people with suspected epilepsy.Added value of this studyWe show that data-driven analysis of routinely collected EEGs that are currently considered clinically non-informative (i.e. absence of apparent epileptiform activity) can be used to distinguish EEGs from people with epilepsy from people with an alternative diagnosis with better-than-chance performance. To the best of our knowledge, this is the largest retrospective study assessing the performance of computational biomarkers derived from clinically non-contributory EEG recordings. The resulting statistical model is interpretable and relies on both spectral and computational (network- and model-based) features. We perform a series of validity and sensitivity analysis to assess the overall robustness of the final statistical model used for classification. We also conduct several statistical tests to analyse any shared characteristics (e.g. site, comorbidity) amongst the primary classes (FP, FN, TP, TN). These findings validate previous biomarker discovery- or development-studies, and provide evidence that they offer better-than-chance performance in a clinically relevant context. Future large-scale studies could consider combining these methods with interictal features for non-specialist settings.Implications of all the available evidenceOur study presents evidence that computational analysis of clinically non-contributory EEGs could provide additional decision support for both epilepsy and alternative conditions. Since the statistical model and underlying features are interpretable, they could provide the starting point for further exploring the mechanisms that drive overall seizure-likelihood. Future work should focus on prospective testing and validation (e.g. identification of specific situations or cases in which these methods could be of added value) as well as assessing heterogeneity across different syndromes and diagnoses (e.g. NEAD, focal vs generalised epilepsy).