Skip to main content
Fig. 3 | Acta Neuropathologica Communications

Fig. 3

From: Machine learning modeling of genome-wide copy number alteration signatures reliably predicts IDH mutational status in adult diffuse glioma

Fig. 3

Cross-validated IDH mutation classifier development. A Cross-validated results showed that IDH mutation logistic regression classifiers trained on chromosome arm resolution SCNA data performed better than classifiers trained on PCA reduced gene-level or PCA reduced cytoband-level SCNA data. B Downsampling gene-level SCNA data to chromosome arm resolution smoothed noisy SCNA data derived from older cytogenetic arrays. C A logistic regression model trained on TCGA SCNA data aligned to hg19 and optimized for maximizing the AUC score performed better than other parameter choices. D Logistic regression mostly outperformed other model classes, including an ensemble of all listed models, across five metrics. E Our model performance increased monotonically when restricted to samples of increasing prediction confidence. This indicated that the calibration of our model’s output probabilities was effective. Standard deviation values for each metric over 1000 cross-validation trials are shaded in. F Restricting model predictions to those made with confidence greater than 0.7 greatly increased model performance

Back to article page