Abstracts

Rationale: Post-traumatic epilepsy (PTE) is one of the common consequences of TBI occurring in patients with brain trauma. Prediction of PTE can help in the development of preventive care for subjects who would be identified as at risk for PTE. Prediction of PTE is very challenging due to the heterogeneous nature of TBI injury types, pathology, and lesions. Alongside structural features such as lesion size and location, functional features obtained from resting fMRI (rfMRI) carry information characterizing the injury. Resting fMRI in TBI subjects is affected through the presence of lesions, both locally and potentially throughout networks in which the lesioned regions are involved. We hypothesize that combining both structural and functional features benefits the prediction of PTE.

Methods: We trained four machine learning methods: (i) Random Forests (RF), (ii) Support Vector Machines (SVM), (iii) a fully connected Neural Network (NN), and (iv) a Graph Convolutional Network (GCN), to predict PTE. We also explored use of principal components analsis (PCA) with each of these methods to reduce the dimensionality of highly correlated features. We use the connectivity matrix (Pearson correlation coefficients) between 16 regions of interest (ROIs) defined in the USCLobes atlas (http://brainsuite.org/usclobes-description) as the functional features for the classifier. This atlas consists of lobar delineations (white matter, left and right frontal, parietal, temporal, and occipital lobes, cingulate gyrus, as well as the bilateral insulae, brainstem, corpus callosum, and cerebellum). We also use lesion volume in 13 ROIs (we removed white matter, brainstem, and corpus callosum) as structural features. Lesion volumes were computed using an automated lesion delineation method based on a variational autoencoder.

Results: Classification results obtained with the machine learning models are reported in Table 1 for lesion volume, connectivity features, and both. KSVM, RF, and NN all showed lower Area under the ROC Curve(AUC) compared to linear SVM, (AUC=0.64, σ= 0.05). Results based on connectivity features were, on average, slightly higher than for lesion features. Using a combination of both connectivity and lesion features increased the AUC to 0.74 (σ= 0.05) for KSVM with PCA, which is significantly larger than either structural or functional features separately. Due to the relatively small size of the training set and the high dimensionality of the features, deep learning (GCN and NN) methods did not improve the results. The methods tested here should be viewed as a baseline and may improve from a different choice of features and, modifications of the network structure and training approach, and a larger set of subjects for training.

Conclusions: Our results show that combining lesion volume data and fMRI based connectivity analysis has potential for predicting PTE onset in TBI patients. Further, a kernel SVM approach appears better suited to this problem than the random forest, SVM, NN and GNN approaches. We also showed using PCA for feature reduction is more effective to prevent over-fitting than using a regularizer.

Funding: Please list any funding that was received in support of this abstract.: R01 NS074980, W81XWH-18-1-0614, R01 NS089212, and R01 EB026299.