Table 1.
Algorithms.
Fig 1.
Meningioma outcomes according to conventional predictive factors.
Kaplan-Meier estimates for local freedom from recurrence (top) and overall survival (bottom) following surgery for meningioma according to conventional predictive factors of meningioma grade (left) and extent of resection (right).
Table 2.
Patients, meningiomas, treatments and outcomes.
Fig 2.
Meningioma feature correlation heatmaps.
Heatmaps based on features’ pairwise Pearson correlation. Rows and columns have been arranged by hierarchical clustering to reveal each feature’s relationship to the outcome of interest. Orange denotes positive correlations, and teal indicates negative correlations.
Fig 3.
Machine learning model accuracy.
(A) Mean balanced accuracy across 100 subsamples of models predicting local failure (top) and overall survival (bottom). Error bars indicate standard deviation. (B) Receiver-operator characteristic curves for local failure (left) and overall survival (right) for the best preoperative (SVM and GLMNET, respectively), conventional (RF and GLMNET, respectively) and integrated models (SVM and GBM, respectively) as defined in (A). 95% confidence intervals were estimated using 2000 bootstraps. All results are reported on left-out test sets not used for model training.
Fig 4.
Variable importance and decision trees.
(A) Mean variable importance derived from 100 random forest models predicting local failure (top) and overall survival (bottom) for preoperative (left), conventional (middle) and integrated (right) models. Error bars indicate standard deviation. (B) Decision trees built using MediBoost Tree-Structured Boosting predicting local failure (top) and overall survival (bottom) corresponding to preoperative (left), conventional (middle) and integrated models (right). N indicates number of meningiomas (LF) and number of patients (OS) that fall into each branch and percentage indicates proportion of those with local failure or deceased, respectively. Unlike tradition decision trees, MediBoost chooses splits based on weighted versions of the full sample at each node, making splits more reliable even as tree depth increases. PCF: posterior cranial fossa; 1, Caucasian; 2, Black; 3, Asian; 4, Hispanic; 5, Pacific Islander; 6, Other; not Hispanic/Latino; 7, White; Hispanic/Latino.
Fig 5.
Machine learning nomograms for meningioma outcomes.
Nomograms built using a penalized Cox model (adaptive elastic net) to predict 5-year freedom from LF (left) and 5-year survival (right) on the full sample. This procedure provides accurate survival estimates even in the presence of correlated features, which are not allowed in the original Cox regression model. Scatter plots show observed versus predicted probabilities obtained by training on 100 bootstrap samples and testing on the left-out set.