Authors :
Presenting Author: Wessel Woldman, PhD – University of Birmingham (UK)
Luke Tait, PhD – Cardiff University; Lydia Staniaszek, PhD – University Hospitals Bristol and Weston NHS Foundation Trust, Bristol, United Kingdom; Elizabeth Galizia, MD – St George's Hospital NHS Foundation Trust, London, United Kingdom; David Martin-lopez, MD – St George's Hospital NHS Foundation Trust, London, United Kingdom; Matthew Walker, MD, PhD – University College London, London, United Kingdom; Al Azeez, MD – University College London Hospitals, London, United Kingdom; Kay Meiklejohn, MSc – University Hospital Southampton NHS Foundation Trust, Southampton, United Kingdom; David Allen, MD – University Hospital Southampton NHS Foundation Trust, Southampton, United Kingdom; Chris Price, MD – Royal Devon and Exeter NHS Foundation Trust, Exeter, United Kingdom; Sophie Georgiou, PhD – Royal Devon and Exeter NHS Foundation Trust, Exeter, United Kingdom; Manny Bagary, MD – Birmingham and Solihull Mental Health NHS Foundation Trust, Birmingham, United Kingdom; Sakh Khalsa, PhD – Birmingham and Solihull Mental Health NHS Foundation Trust, Birmingham, United Kingdom; Francesco Manfredonia, MD – The Royal Wolverhampton NHS Trust, Wolverhampton, United Kingdom; Phil Tittensor, MD – The Royal Wolverhampton NHS Trust, Wolverhampton, United Kingdom; Charlotte Lawthom, MD, PhD – Swansea University, Swansea, United Kingdom; Rohit Shankar, MD, PhD – University of Plymouth, Plymouth, United Kingdom; John Terry, PhD – University of Birmingham, Birmingham, United Kingdom
Rationale:
A routine electroencephalography (EEG) is a common investigation for suspected epilepsy cases. Confirming diagnosis with EEG requires the presence of interictal epileptiform discharges (IEDs), and convincing clinical history concordant with IED appearance and location. However, IEDs are not always present in EEGs from persons with epilepsy; routine EEG has a sensitivity of 17-56%. As EEG can only be used to include a diagnosis of epilepsy (not its exclusion), EEGs lacking IEDs do not contribute either for or against a diagnosis of epilepsy.
To improve sensitivity of routine EEG, we developed a set of candidate biomarkers of seizure susceptibility on portions of EEG lacking IEDs. Computer modeling of dynamic networks have previously shown sensitivity and specificity in datasets comparing people with idiopathic generalized epilepsy to healthy controls and in small-scale, case-controlled studies using controls with alternate (to epilepsy) neurological conditions. The objective of this study was to validate the biomarkers in non-contributory EEGs from a large, representative cohort of patients presenting to various diagnostic centres with suspected epilepsy.
Methods:
This retrospective, multi-site, case-controlled study used non-contributory EEG from patients who ultimately received a stable (>1 year) diagnosis of epilepsy or a common alternate condition (e.g., syncope, functional seizures). Non-contributory EEGs (n=290) were collected from eight UK medical centers and classified as normal (n=197) or abnormal (containing non-specific abnormalities not intrinsically indicative of epilepsy, n=93) by metadata supplied by sites. Eight candidate markers (spectral [n=2], network-based [n=4], model-based [n=2]) were calculated within each recording. A classifier was developed using ensemble-based learning methods with a two-tier cross-validation approach. Standard regression models were used to identify if potential confounding variables (e.g., age, gender, treatment-status, comorbidity) correlated with performance.Results:
A clinical diagnosis of epilepsy was consistent with higher model scores in both normal and abnormal non-contributory EEGs (see Figure 1 and Table 1). We observed balanced accuracies of 67.9% (sens: 61.2%, spec: 74.6%) in normal non-contributory EEGs and 61.15% (sens: 61.4%, spec: 60.9%) in abnormal non-contributory EEGs. Group-level analysis found no evidence suggesting any of the potential confounding variables significantly impacted the overall performance.Conclusions:
Our candidate set of biomarkers showed promising levels of sensitivity and specificity in this first-of-its-kind large, multi-site cohort of non-contributory EEGs. In contrast, none of the considered EEGs were used to contribute for or against a diagnosis of epilepsy in the clinical setting. Our findings provide evidence that the set of biomarkers could improve diagnostic accuracy of routine EEG.Funding:
This work was supported by the NIHR (AI01646), Epilepsy Research UK (F2002), Innovate UK (103939), EPSRC (EP/N014391/2 & EP/T027703/1).