Table 5.
Machine-learning selected parameters for diagnostic discrimination of responders and non-responders.
| Computationally selected features for the multi-centric clinical trial data subset (0–180 days) N = 58 |
Weights for the selected features | Computationally selected features for the clinical trial data and laboratory biomarker subset of the Rostock group (day 0 - preoperative) N = 31 |
Weights for the selected features |
|---|---|---|---|
| DeltaViable tissue 6 m/0 | 2.554 | NT proBNP 0 | 9.718 |
| Triglycerides 0 |
2.260 | VEGF_I | 7.810 |
| Scarsize 6 months |
2.159 | Erythropoietin_I | 4.262 |
| DeltaScarsize 6 m/0 |
2.063 | Vitronectin_I | 3.898 |
| Nonviable tissue 6 months |
1.999 | CFU_Hill_I | 2.871 |
| Body mass index 0 |
1.982 | CD45Neg_EPC_I | 2.186 |
| 6MWT 0 |
1.974 | CD117_184_PB_EPC_IHG_I | 2.146 |
| DeltaEF 6 m/0 |
1.967 | CD45_117_184_EPC_I | 2.118 |
| 6MWT 10 days |
1.920 | CD45_133_146_PB_CEC_I | 1.969 |
| LVEF 0 |
1.890 | Thrombocytes I | 1.951 |
| Bypasstime min | 1.883 | IGFBP-3_I | 1.922 |
| Euroscore 0 |
1.874 | CD133 pro ml PB_I IHG | 1.910 |
| CKmax | 1.857 | CD146_PB_CEC_I | 1.799 |
| Scarsize 0 |
1.771 | CD105_PB_CEC_I | 1.793 |
| NTproBNP 0 |
1.771 | CD45_133_34_105_PB_CEC_I | 1.489 |
| Crossclamptime | 1.675 | MatrigelPlug_PB_31_I | 1.475 |
| Delta6MWT 6 m/0 |
1.673 | CD45_133_34_117_309_EPC_I | 1.420 |
| Creatinine 0 |
1.645 | Delta_CT_SH2B3_I | 1.393 |
| LVESV 0 |
1.604 | Weight | 1.363 |
| Weight 0 |
1.389 | LVESV I 0 | 1.352 |
| Accuracy | 63.35% | Accuracy | 81.64% |
Selected features of the AdaBoost ML algorithm showing the most informative selection criteria for the subsequently created ML models. The features are ordered due to their calculated weights in a decreasing manner. Accuracies are based on 100 independent predictions of 10-fold cross-validation calculations (Model has been built after AdaBoost feature selection and random forest feature learning).