Extended Data Fig. 2: Model comparisons for best prediction accuracy.
From: Genomic copy number predicts esophageal cancer years before transformation
a, Shows the comparison of the model used in the analysis presented (trained on all samples, n = 773) versus a model which excludes the most extreme histopathological samples (excluding HGD/IMC, n = 711). We compare the accuracy of the ROC AUC using the best sensitivity threshold (Pr = 0.3) presented in Fig. 2a of the main paper. A model trained without use of the extreme samples shows no decrease prediction accuracy indicating that these samples are not driving the differences in the model. b, ROC AUC values describing the prediction accuracy for models trained on different sets of data and various aggregations of per-sample predictions also using the best sensitivity threshold (Pr = 0.3). The first set of bars provides the ROC values for the reference model per-sample predictions (n = 773). The following bars describe the ROC values for aggregated predictions on the same samples: mean and max prediction per endoscopy, mean and max prediction per patient (excluding the final HGD/IMC samples). The aggregated predictions do not differ from the per-sample predictions indicating that a single sample may be sufficient for accurate prediction. All error bars denote the 95% confidence interval for the sensitivity, specificity, and AUC at a threshold of Pr = 0.3.
