Abstract
PURPOSE: Seizures are frequent symptoms of gliomas. Predicting which patients are more likely to seize and in turn require anti-epileptic management has been a challenge. Correctly identifying these patients could help optimize care and minimize side effects. To provide guidance for clinicians, we used machine learning techniques to predict seizure presentation in this population.
METHODS
We used volumetric data of pre-treatment MR images (T1Gd and T2-FLAIR sequences), patient demographics (age; sex), and measurements of tumor proliferation (log()), invasiveness (log(D)) and their relative ratio (log(/D)). We compared the performance of 5 machine learning models in predicting seizure status, using Artificial Neural Network, Naive Bayes (NB), Linear Discriminant Analysis (LDA), Random Forest, and Support Vector Machine. Correlations between probability of seizure presentation (p(SP)) and continuous variables were also analyzed.
RESULTS
Our cohort consisted of 59 seizure-presenting and 77 non-seizure-presenting patients. All models consistently demonstrated significant correlations (p < 0.05) between (p(SP)) and the following variables: T1Gd radius (-0.781 to -0.674), T2-FLAIR (-0.674 to -0.611), and log(/D) (0.169 to 0.294). Age was significant (p < 0.05) in 4 of the 5 models (-0.211 to -0.175). Mean performance measures for the models (and the best performer) were: 0.726 for Area under the ROC curve (0.75 with NB), 0.6202 for sensitivity (0.661 with NB), 0.74 for specificity (0.766 with LDA). The 5 features ranked as most important were: T1Gd, T2-FLAIR, log(/D), age, log(). CONCLUSIONS: We found an association in seizure-presenting patients with smaller, more proliferative tumors and younger age. Machine learning predictive modeling can potentially be informative in the clinical arena. Further validation studies, to determine the degree of data overfitting, model versatility, as well as performance on test data, are warranted.