Fig. 2: Model performance and variable importance.

a Receiver operating characteristics (ROC) curve illustrating the tradeoff between model sensitivity and specificity based on mode-predicted probabilities of TSRD in the independent test sample. b Logistic calibration curve illustrating the correspondence between model-predicted probabilities and observed prevalence of TSRD in the independent test sample. c Shapley Additive Explanations (SHAP) plot illustrating the impact of each predictor on model predictions in the development sample. Predictors are arranged on the y-axis in order of absolute mean contribution, with exact values provided next to each predictor. Positive or negative SHAP values on the x-axis indicate higher or lower predicted probability of TSRD, respectively. In the plot, each point represents a participant in the development sample, and the color represents the value of the predictors. For example, higher values of age (blue) yielded lower predicted probabilities of TSRD, and vice versa.