Lateralization of temporal lobe epilepsy by multimodal multinomial hippocampal response-driven models

Abstract

Purpose Multiple modalities are used in determining laterality in mesial temporal lobe epilepsy (mTLE). It is unclear how much different imaging modalities should be weighted in decision-making. The purpose of this study is to develop response-driven multimodal multinomial models for lateralization of epileptogenicity in mTLE patients based upon imaging features in order to maximize the accuracy of noninvasive studies. Methods and Materials The volumes, means and standard deviations of FLAIR intensity and means of normalized ictal–interictal SPECT intensity of the left and right hippocampi were extracted from preoperative images of a retrospective cohort of 45 mTLE patients with Engel class I surgical outcomes, as well as images of a cohort of 20 control, nonepileptic subjects. Using multinomial logistic function regression, the parameters of various univariate and multivariate models were estimated. Based on the Bayesian model averaging (BMA) theorem, response models were developed as compositions of independent univariate models. Results A BMA model composed of posterior probabilities of univariate response models of hippocampal volumes, means and standard deviations of FLAIR intensity, and means of SPECT intensity with the estimated weighting coefficients of 0.28, 0.32, 0.09, and 0.31, respectively, as well as a multivariate response model incorporating all mentioned attributes, demonstrated complete reliability by achieving a probability of detection of one with no false alarms to establish proper laterality in all mTLE patients. Conclusion The proposed multinomial multivariate response-driven model provides a reliable lateralization of mesial temporal epileptogenicity including those patients who require phase II assessment.

MRI characterization of temporal lobe epilepsy using rapidly measurable spatial indices with hemisphere asymmetries and gender features

“…The quantitative ICC in reliability testing [56,58,65,66] and uncertainty analysis by Bayesian inference based on the posterior probability distribution [20,[67][68][69] are used for precise interpretation of MRI data and prediction of dynamic structural/functional model analysis. In the present study, the variance among healthy subjects, i.e., within-subject variability and the between-subject variability, had been assessed by the reliability testing through estimation of the ICC with 95 % confidence intervals using one-way ANOVA.…”

“…The ample evidences of advanced research using multimodal combinational non-invasive neuroimaging techniques like MRI, MRS, f-MRI, FDG-PET, SPECT, and MEG have revealed that most of the cases of TLE patients are accompanied by pathological changes in brain structure and function network, including cortical and subcortical (hippocampus and parahippocampus) structures [12][13][14]. Such multimodal combinational imaging (structural and functional) approaches are being suggested to detect the laterization of TLE with ictal and interictal onset in pre-surgical diagnostic purpose [15][16][17][18][19][20]. Regardless, epilepsy occasionally is a clinical problem as regards medical management and the problem becomes acute in developing countries as there is often delay in clinical diagnosis due to the limited resources, large patient load, and health care manpower constraints [21][22][23].…”

MEG Coherence and DTI Connectivity in mTLE

“…For both coherence and nodal degree, statistically significant laterality measures between the left and right mTLE cohorts were considered as multivariate independent variables and incorporated into the development of response-driven models of laterality using logistic function regression (Hosmer Jr et al, 2013, Nazem-Zadeh et al, 2014a, Nazem-Zadeh et al, 2014b). …”

Application of DTI connectivity in lateralization of mTLE

“…In order to assess how the multinomial logistic function generalized to an independent dataset and how accurately this response model performed in practice, a cross-validation was performed using the ‘leave-one-out’ approach [16]. The probability of detection on the test data was then calculated [17, 18]. …”