Fig. 3: Performance of GAM–LASSO MPR predictive models.
a, Use of the GAM–LASSO model to predict MPR in test set 2, which consisted of patients within LCMC3 who were not included in either the training set (n = 57) or test set 1 (n = 54). MPR was not assessed in these nine patients because of no resection. The MPR and non-MPR cohorts derived from the merge of the model’s training set and test set 1. The maximum and minimum values of the boxes denote the IQR. The line within the IQR denotes the median. The extremities of the dashed lines represent the minimum and maximum values of the data, which are 1.5× below the first quartile and 1.5× above the third quartile. The parameters for null hypothesis testing via analysis of variance (ANOVA) were as follows: d.f. = 2, total sum of squares = 1.976, mean squares = 0.988, F-value = 32.799 and Pr(>F) = 4.914 × 10−12. The statistical details for the comparison of MPR and non-MPR were t = −5.47, d.f. = 27.02, two-sided P = 8.6 × 10−6 and 95% CI = −0.439 to −0.200. The statistical details for the comparison of MPR and PD were t = −3.18, d.f. = 28.45, two-sided P = 0.0035 and 95% CI = −0.383 to −0.083. The statistical details for the comparison of non-MPR and PD were t = −1.77, d.f. = 9.52, two-sided P = 0.11 and 95% CI = −0.195 to 0.023. No adjustment was made for multiplicity. b, ROC curves for the training set and test set 1. The dashed y = x line, which represents random assignment, is included for reference. aImmunophenotyping via flow cytometry. IQR, interquartile range; ROC, receiver operating characteristic.
