Joeky Senders

98 Chapter 5 TABLE 3. Secondary metrics for model performance and deployment. Model Interpretability Predictive Applicability Computational Efficiency a Inference Prediction Binary Continuous Survival Curves Size (Mb) Load Time (s) Prediction Time (s) AFT X X X X X 20 0.9 1.9 Bagging - X X X - 16,380 1,335 31.8 Blackboost - X X X X 36,790 2,455 234.3 CPHR X X X X X 37 1.7 7.5 Recursive Partitioning - X X X X 490 52.1 3.4 BDT - X X X - 300 8.2 2.1 GLM X X X X - 1 0.2 1.7 GLMnet X X X X - 109 6.7 2.3 K-Nearest Neighbors - X X X - 91 5.6 1.9 Multilayer Perceptron - X X X - 45 1.4 17.4 Naïve Bayes - X X - - 82 2.9 13.0 Random forest - X X X - 1,100 41.4 10.1 Random Forest Survival - X X X X 6,350 65.7 139.0 Support Vector Machine - X X X - 111 4.8 4.4 XBDT - X X X - 92 2.4 1.5 Abbreviations: AFT=accelerated failure time; CPHR=Cox proportional hazards regression; GLM(net)= (Lasso and elastic- net regularized) generalized linear models; Mb = megabyte; s = seconds; TTE=time-to-event; (X)BDT= (extreme) boosted decision trees a Based on a 100-fold bootstrapped model. Deployment Although the AFT model demonstrated similar to superior performance in terms of discrimination and calibration, it outperformed competing statistical and machine learning algorithms in terms of interpretability, predictive applicability, and computational efficiency. Therefore, it was selected as back end for the online survival prediction tool. (https://cnoc-bwh.shinyapps.io/gbmsurvivalpredictor/ ). The estimated survival profile for a hypothetical patient is shown in Figure 3.