Fig. 1

Model performance assessed by sensitivity at 98.6% specificity (panel A) and area under the receiver operating characteristic curve (panel B). The diagnosis-only rule-based method involves no parameter tuning, so it is reported for all participants. All other models are analysed by train and test population. Training results are calculated using cross-fold hold-outs and are plotted with mean and 95% confidence interval for standard error. Testing results are calculated using n = 2000 bootstrapped samples and are plotted by empiric 95% confidence interval. Area under the curve is not calculable for the rule-based diagnosis-only method