Fig. 5

Developing an optimal clinical and multi-‘omic model of rapid vs. late relapse in TNBC. a Schematic of experimental steps including definition of variables, descriptive statistics, comparative modeling including model tuning, and assessment of model performance. b Receiver-operator characteristic (ROC) plots for each model’s performance, measured by average area under the curve (AUC) of 25 independent runs of the train-test split. Each model was tuned to ensure optimal performance. Models are grouped and colored by cohort—red indicates training data (n = 63), green indicates testing data (n = 63), and blue indicates the independent validation Fudan cohort (n = 34). For each grouping, the three models shown are: 1) “null model”, including only clinical variables; 2) “null plus significant features”, adding any feature significantly different between rrTNBC and lrTNBC with a nominal p-value < 0.05; and 3) “null plus significant features reduced”, including only features from model 2 that are among the top 25 most important genes in at least half of the independent runs. Asterisks indicate significance by Wilcoxon rank sum, * indicates p < 0.05, ** indicates p < 0.01, NS indicates “not significant” (p > 0.05)