FIGURE 2 from Multi-institutional Prognostic Modeling in Head and Neck Cancer: Evaluating Impact and Generalizability of Deep Learning and Radiomics

<p>Model results. <b>A,</b> Overview of model characteristics. The characteristics are grouped into input data, prediction head (i.e., how were the survival predictions made) and model type (whether the model involved any nonlinearities and/or convolutions). PH: proportional hazards, MLP: multilayer perceptron, *: age, sex, stage, HPV status. <b>B–D</b>, Performance of all models, including benchmark models, in terms of 2-year AUROC, 2-year average precision and <i>C</i>-index of the lifetime risk, respectively. The results are ranked by AUROC (numbers above bars indicate the overall rank of each model). Error bars represent 95% confidence intervals computed using 10,000 stratified bootstrap replicates. Dashed gray lines indicate random guessing performance (0.5 for AUROC and <i>C</i>-index, 0.14 for AP). <b>E–G</b>, show the Kaplan–Meier survival estimates in low- and high-risk groups identified by the best performing model in each category (combined, EMR only and radiomics), respectively. Test set patients were stratified into two groups based on the predicted 2-year event probability at 0.5 threshold. In each case, there were significant differences in survival between the predicted risk groups (HR, 8.64, 5.96, and 4.50, respectively, <i>p</i> < 10₋₁₈ for all).</p>