Parameters were selected as follows: if a parameter proved to be significant in distinguishing between the two groups (Table 1) and it also proved to contribute to this distinction when combined with other parameters (Table 2, showing the ranking obtained for all possible model configurations), it was selected to be one of the parameters of the optimal sub-set. The respective receiver operating characteristic (ROC) curves for each parameter are depicted in Figure 2 (top: the six selected parameters; bottom: the eight neglected parameters). Of the top five performing measures in Table 2, four (namely, C, k_0.95, NMSE₃, and MSE) were also significant measures individually (Table 1). This suggests that they are good candidates for inclusion in an optimal sub-set of AF substrate complexity measures. fWPM was not significant when considered individually, but ranked first when analysing all possible model configurations. This may mean that it likely carries independent information which helps discriminating a few subjects among the two groups. This measure can be seen as an alternative to fibrillatory wave amplitude, known to be associated with the degree of AF organization and outcome in patients undergoing catheter ablation for AF.¹⁶ For these reasons, it was included in the optimal sub-set. fWP was ranked sixth, and it was excluded due to its high correlation with fWPM (Spearman correlation coefficient r = 0.68, P < 0.01). SAE was also selected, because it measures whether recurrent patterns are present in the data, hence focusing on different characteristics of the data than the parameters that were already selected. MOI and RHE were excluded due to both of their low individual significance (0.04 < P < 0.05) and their low ranking. All remaining parameters were excluded. The optimal sub-set attained a perfect classification when used as input to a multivariable regression model.