Authors :
Presenting Author: Soumen Ghosh, PhD (Submitted), MTech – The University of Queensland
Viktor Vegh, PhD – Associate Professor, Centre for Advanced Imaging, The University of Queensland, Australia; Allice-Ann Sullivan, MBBS – The Royal Brisbane and Women's Hospital; Shahrzad Moinian, PhD – Postdoctoral Research Fellow, Centre for Advanced Imaging, The University of Queensland, Australia; John Phamnguyen, MBBS, FRACP (Neurology) – PhD Student, Centre for Advanced Imaging, The University of Queensland, Australia; David Reutens, MBBS, MD, FRACP, FAHMS – Emeritus Professor, Centre for Advanced Imaging, The University of Queensland, Australia
Rationale:
Seizure recurrence is a significant concern for individuals after their first unprovoked seizure. Untreated, up to 50% of patients develop epilepsy within two years of the initial seizure. Despite the absence of currently-used predictors, approximately 30% of first-seizure patients experience a second seizure. Predicting seizure recurrence is challenging, especially in patients with otherwise unremarkable clinical history and normal MRI scans. Prophylactic antiseizure medications are also prescribed in individuals who would not have had recurrent seizures without treatment. Thus, new approaches for predicting seizure recurrence after the first unprovoked seizure are required.Methods:
We propose a machine learning framework (Figure 1) to predict the chance of seizure recurrence for patients with the first unprovoked seizure by utilizing routine clinical 3T brain MRI scans and clinical information available to clinicians in the first seizure clinic. Our study cohort consisted of 169 subjects categorized into individuals with (n=24) and without (n=145) seizure recurrence. Brain MRI radiomic features were constructed from FreeSurfer statistics are fused with coded clinical information to build the Machine Learning (ML) model. To address the issue of class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) algorithm was applied to balance both groups by upsampling the minority group. Finally, six ML classifiers were built using the feature subset selected by a K-nearest neighbors-based sequential feature selection method. The classifiers were trained using 70% of the dataset and model performance tested on the remaining 30%. Results:
One hundred sixty nine individuals (102 males, 67 females) with first unprovoked seizure were included in this study. The average age at the time of the first seizure was 37.7 years (standard deviation 16.3 years). The ML model performed well in predicting seizure recurrence using 12 features (five cortical, three subcortical, and four clinical features, see Table 1). The Extreme Gradient Boosting (XGB) classifier outperformed other classifiers with an area under the curve, classification accuracy, sensitivity, specificity, and Dice score of 96%, 93%, 91%, 95%, and 93% respectively. The lowest performance was achieved by the Decision Tree (DT) classifier for which the corresponding results were 82%, 82%, 79%, 84%, and 81% respectively.Conclusions:
Our machine learning framework, which uses hand-crafted brain MRI features and clinical information, may enhance clinical decision-making in relation to seizure recurrence after the first unprovoked seizure. The brain feature metrics and clinical information highlight the potential brain differences in individuals with recurrent seizures. Our study underscores the potential of using data-driven techniques to improve the accuracy of predicting seizure recurrence.Funding:
This work is supported by Australian Research Council funded Training Centre for Innovation in Biomedical Imaging Technology (CIBIT; IC170100035) and Royal Brisbane Women’s Hospital (Brisbane, Australia) Foundation grant both awarded in 2018.