Fig. 5.

a Three-dimensional t-SNE calculation of the Rostock subgroup. The variables x and y refer to the newly calculated features that are used to classify the patients into distinct groups. The model was subsequently fitted by a polynomial (n³) equation to visualize the z-axis as a geographic profile. The respective colors for the responder (red dot) and non-responder (grey dot) patients have been added afterwards. The classified groups have been roughly summarized by a red and grey dashed line. Results are obtained after 3000 iterations. The calculation of the ratio between responder and non-responder is indicated for each circle. It is more likely for the non-responder group to be located at smaller z-values (z < 20, ratio < 42%). The responders tend to be enriched within the light blue areas (z > 20) including a ration > 69%. b Obtained supervised ML prediction results for pre- and postoperative time points (0 days to 180 days) of the clinical and clinical & laboratory dataset to distinguish between responder and non-responder. The graph shows the true positive prediction results of five independent feature selected ML models (AdaBoost for feature selection and RF for final prediction).The error bars indicate the respective accuracy standard deviation for the constructed models that have been obtained after 100 iterations. The 100 model iterations are significant different according to one-way ANOVA (p < 0.001).