Prediction of Post Traumatic Epilepsy Using MR-based Imaging Markers
Abstract number :
2.204
Submission category :
5. Neuro Imaging / 5A. Structural Imaging
Year :
2023
Submission ID :
649
Source :
www.aesnet.org
Presentation date :
12/3/2023 12:00:00 AM
Published date :
Authors :
First Author: Anand Joshi, PhD – University of Southern California
Presenting Author: Wenhui Cui, – University of Southern California
haleh Akrami, PhD – University of Southern California; Andrei Irimia, PhD – University of Southern California; Wenhui Cui, PhD – University of Southern California; Christianne Heck, MD PhD – University of Southern California; Paul Kim, MD – University of Southern California; Karim Jerbi, PhD – University of Southern California; Dileep Nair, MD – Cleveland Clinic; Richard Leahy, PhD – University of Southern California
Rationale: Post-traumatic Epilepsy (PTE) is a debilitating neurological disorder that develops after traumatic brain injury (TBI). Despite the high prevalence of PTE, current methods for predicting its occurrence remain limited. In this study, we aimed to identify MRI imaging-based markers for the prediction of PTE using machine learning.
Methods: We extracted functional and structural features from two datasets: The Maryland TBI MagNeTs dataset (Gullapalli 2011) (N=113 total), 72 (36 PTE and 36 non-PTE groups) were used for group-level difference comparisons as well as for supervised machine learning (i.e., constructing a model from training samples to predict the presence of PTE in previously unseen data). The remaining 41 non-PTE subjects were used to train an artificial neural network for automatic lesion delineation. The TRACK-TBI Pilot dataset (Yue 2013) (N=97) subjects were used to train the artificial neural network for the automated delineation of lesions. The data was preprocessed using tools from BrainSuite software and coregistered to USCBrain atlas for statistical group comparisons. We examined three imaging features: lesion volumes, resting-state fMRI-based measures of functional connectivity and amplitude of low-frequency fluctuation (ALFF). We employed four machine learning methods, namely, support vector machines (SVM), kernel support vector machine (KSVM), random forest, and a neural network, to develop predictive models. Furthermore, we performed voxel-wise and lobe-wise group difference analyses to investigate the specific brain regions and features that the model found to be most helpful in distinguishing PTE from non-PTE populations. The group comparisons were performed using F-test with permutations with FDR correction for multiple comparison correction. The feature importance was computed using SVM weights for features.
Results: Our results showed that the KSVM classifier, with all three feature types as input, achieved the best prediction accuracy of 0.78 AUC (Area Under the Receiver Operating Characteristic (ROC) curve) using nested cross-validation. Our statistical analysis of feature importance using F-tests with permutations, as well as SVM-based feature importance maps, show converging evidence of significant differences in bilateral temporal lobes and cerebellum between PTE and non-PTE groups.
Conclusions: Our findings demonstrate the complementary prognostic value of MR-based markers in PTE prediction and provide new insights into the underlying structural and functional alterations associated with PTE. The best results were obtained with KSVM, which is possibly partly due to the heterogeneity of the alterations in the PTE group around the mean feature. Our results using kernel-based methods show promising results. In lesion and ALFF comparison studies, bilateral temporal lobes and cerebellum show significance. Additionally, parietal lobes might have involvement.
Funding: This work is supported by Department of Defense grants HT94252310149, W81XWH-18-1-061 and by National Institutes of Health grants R01 NS074980, and R01 EB026299.
Neuro Imaging