Abstracts

Rationale: Non-invasive language mapping for presurgical evaluation of hemisphere dominance is often an essential part of a phase-1 workup in refractory epilepsy. While methods for language laterality determination using magnetoencephalography have been well vetted, there remains large variation across labs in terms of methods and still a great deal of reliance on the interpretation of the ECD model fit or cortical response function. With the development of learning algorithms for decoding subtle state changes in MEG, classification methods may provide a model free method for language laterality determination. Here we test the hypothesis that classification based on decoding word vs. tone stimuli separately in left and right MEG sensors will result in higher significant classification accuracy in the language dominant hemisphere.   Methods: MEG data (Elekta Triux, Elekta Oy., Ltd.) from four refractory epilepsy patients who had undergone presurgical language and auditory functional mapping were used.  WADA, functional MRI and/or surgical outcome were used as confirmatory evidence of hemispheric dominance.  The data were cleaned using the tsss spatial filter provided by the MEG manufacturer and trials containing artifact or epilepsy discharges removed. We tested a linear support vector machine algorithm (SVM) on the data filtered in 3 frequency bands and with or without prior common spatial filter decomposition (CSP). Data representing the left and right sensors were classified independently.  For decoding accuracy, we applied a standard cross-validation method for prediction of classification for the SVM on individual time points and a monte-carlo cross validation for prediction on the computed CSP data collapsed across the epoch. The computed results were the scores of the estimates across time, comparing the prediction estimated for each trial to its true value.   All data analysis was performed within the MNE-python MEG analysis toolbox.  Results: For the SVM applied to the time-point data, ttests over all time points provided significant indices of laterality (p < 0.001) when comparing left vs. right hemisphere sensors for decoding accuracy. The laterality in these cases was consistent with the laterality determined by auxiliary confirmatory means, including a WADA determined right hemisphere dominance in patient 2. The CSP data also provided consistent positive prediction for laterality although statistical methods for this approach are still being evaluated. Overall, using the largest bandwidth (2-90Hz) provided the most significant results.  Conclusions: Our preliminary results suggest decoding methods may provide a useful model free method for determination of hemisphere language dominance. Further work is needed to optimize the pipeline and improve robustness of this and other potential algorithms, as well as validate this approach in a large cohort of patients.  Funding: none