The training sets and the validation sets were comparable regarding patients’ characteristics in terms of the chosen variables, as displayed in Table 2 for the witnessed cases and (supplementary data) for the non-witnessed ones. The best prediction results could be achieved with simple linear logistic regression. The PI3K inhibitor average AUC (the mean over 10 cross-validation runs, as measured for each

of the 10 hold-out sets) was 0.827 (CI 0.793–0.861) for witnessed out-of-hospital cardiac arrest and 0.713 (CI 0.587–0.838) for non-witnessed out-of-hospital cardiac arrest. These results were consistently better (or at least were not significantly worse) than any neural network (multilayer perceptions with 2–10 hidden units) that was tested (results not shown). Therefore, all of the remaining results reported concern the linear predictor version. For witnessed out-of-hospital Akt inhibitor cardiac arrests, the total AUC using all of the variables was significantly better than the AUC for any single considered variable (see Fig. 2). This was not the case for non-witnessed

out-of-hospital cardiac arrests, for which prediction on the single variable adrenaline achieved a mean AUC of 0.780 (CI 0.650–0.802) (see supplementary files). Although this appears to be better than prediction using all of the variables (see Fig. 2B), it was not significantly different based on a two-tailed paired t-test with α = 0.05. For witnessed out-of-hospital cardiac arrests, the single most predictive variable was also adrenaline (AUC: 0.730 [CI 0.689–0.772]).

The right side of Figs. 2 depicts a ranking of the variables by absolute value of regression coefficient for witnessed and non-witnessed out-of-hospital cardiac arrest patients, respectively. For witnessed cases, the ranking Neratinib in vivo of the variables in Fig. 2A was generated by selectively including variables one by one, beginning with the variable with the highest regression coefficient, minutes to SROSC, and comparing the performances of the limited number of variable predictors with the prediction based on all of the variables. The result is shown in Fig. 2B. For non-witnessed cases, the same type of analysis was not possible, given that the single most predictive variable, adrenaline, was not statistically inferior to using all of the variables. From Fig. 2B, it can be concluded that using the four variables min2srosc, age, shockable, and adrenaline, one achieves practically the same prediction performance as using all 21 variables. Therefore, these four variables were used to devise the main prediction score. There were two ways of carrying out this step. The first was to use the final logistic equation—after repeating model estimation with an extended training set of n = 1095 and the reversal of the original normalisation of each variable—to yield the following values of the regression equation(1) Y=0.0284 × min2srosc+0.0355 × age−1.