Authors :
Presenting Author: Jonah Moss, MD – Rush University Medical Center
Brendan Colgan, BA – Rush University Medical College
Adriana Bermeo-Ovalle, MD – Rush University Medical Center
Travis Stoub, PhD – Rush University Medical Center
Rebecca O'Dwyer, MD – Rush University Medical Center
Rationale:
Functional Seizure Disorder (FSD) is a condition which involves events of paroxysmal changes in cognition, motor function, and behaviors which may mimic epileptic seizures but are without EEG correlate. Although numerous studies have suggested anatomic/radiographic features correlate with FSD, there is no consensus on underlying radiographic markers in FSD. Thus, due to its elusive etiology and its ability to mimic the life threatening medical condition of epilepsy, diagnosis of FSD is a difficult task which consumes healthcare resources, increases patient and system costs, and frequently results in inappropriate treatment with ASDs. This study aims to develop a machine learning model, through a novel un-supervised learning approach, that could aid in the early detection of FSD and increase diagnostic certainty using radiographic data.
Methods:
We identified 66 subjects with a diagnosis of FSD via video-EEG capture of their known spells (without electrographic correlate) and 22 control subjects with a diagnosis of temporal lobe epilepsy (TLE) via retrospective chart review of video-EEG monitoring and collected magnetic resonance imaging (MRI) for each. We divided these MRIs into regional segments using the Freesurfer Software Suite and subsequently generated combinations of MRI regions to be fed into a convolution neural network. Of the 50,000+ models generated by this neural net, approximately 200 models with strong model sensitivity/specificity were selected. These strong models were re-trained using transfer learning from publicly available MRI datasets (MedicalNet) and regionally labeled FSD MRIs.
Results:
Six specific brain regions (henceforth referred to as SA-regions) have been found to be highly associated with well performing models. The SA-regions identified were the bilateral amygdala, cerebellar cortex, and cerebellar white matter. When the base SA-region model was modified to include additional brain regions, varying levels of sensitivity and specificity were observed with different brain region combinations. The current preliminary model (prior to completion of transfer learning) identifies FSD patients via full MRI with ~63% sensitivity and via regionally processed MRI with ~88% sensitivity when compared to patients with TLE .
Conclusions:
Our machine learning model shows promise in its ability to classify FSD or TLE via MRI alone. Furthermore, the transfer-learning process may minimize some of the previously stated limitations/biases, while concurrently strengthening the model’s overall performance. In addition, our model’s identification of strongly associated brain regions provides guidance for future investigations beyond machine learning. Thus, this study represents a notable step in advancing our understanding of FSD via both the machine learning model’s identification of strongly associated brain regions and the model’s ability to differentiate MRIs of FSD subjects from those with TLE.
Funding: None