Introduction: Application of certain machine-learning algorithms have the advantage of constructing more accurate statistical models with smaller data sets, such as those often presented in the neurosurgical literature, compared to traditional statistical methods. Thus, our aim was to investigate applicable machine-learning algorithms to predict meningioma grade based on pre-operative radiographic findings on magnetic resonance imaging.
Background: Prognostication and surgical planning for grade WHO I vs. WHO II meningioma requires thoughtful decision-making based on radiographic evidence, among other factors. While conventional statistical models such as logistic regression are useful, machine-learning algorithms are often more predictive, have higher discriminative ability, and can learn from new data. Herein, we analyze conventional statistical models and machine-learning algorithms to predict atypical meningioma based on pre-operative MRI findings.
Methods: Our cohort included patients age 18-65 with WHO grade I (n=94) and II (n=34) meningioma who had a pre-operative MRI between 1998 to 2010. We trained train several binary classifiers: k-Nearest Neighbors models, Scaled Vector Machines, Naïve Bayesian Classifiers, Artificial Neural Network (ANN), as well as logistic regression models to predict tumor grade. Area Under the Receiver Operator Characteristic Curves (AUROCCs) was used for comparison across and within model classes.
Results: We included six pre-operative imaging and demographic variables: tumor volume, degree of peri-tumoral edema, presence of necrosis, tumor location, presence of a draining vein and gender to construct our models. The ANN outperformed all other models across the true positive versus false positive (ROC) space (AUC=0.8995). Optimized logistic regression with quadratic terms (AUC=0.8423, p=0.002) outperformed logistic regression with interactions (AUC=0.7641, p=0.009). Course Gaussian SVM was the optimal SVM identified by hyperspace parameterization and an AUC of 0.8611, while outperforming all KNNs in 5-fold cross validation (0.2547 versus 0.2738). Naïve Bayesian (AUC=0.7096) and logistic regression with main effect terms (AUC=0.7228) were the least accurate binary predictors.
Conclusions: Machine-learning algorithms are powerful computational tools that can predict meningioma grade with great accuracy, thus potentially greatly improving outcomes for patients with meningioma. It could be envisioned that adoption of these models into an electronic health record would provide the clinician empirical data in real-time, affording an unparelleled level of sophistication in offering treatment recommendations.
MEDICAL IMAGING & BIOMEDICAL DIAGNOSTICS
Author: Andrew Hale
Coauthor(s): David P. Stonko, Megan. K Strother, Lola Chambless
Status: Completed Work