Table 5.
Dworak response grade prediction analysis. Clinical, biological, and genetic predictors filtering was based on the results of five algorithms which measure the contribution to prediction on the grade of tumor regression (Dworak grade) after RCT was administrated prior to surgery (CEA serum levels, lymph node involvement, and alterations on 4q, 15q11.1, 15q14, and 17q21.31 chromosomal regions).
| Algorithm | Parameters | Filtering Method | Nº of Variables | Hit Rate (%) | |||
|---|---|---|---|---|---|---|---|
| G0/G1 | G2 | G3/G4 | Global | ||||
| PLS | Number of factors: 3 | No | 7 | 40 | 100 | 67 | 60 |
| PLS | Number of factors: 2 | Yes | 4 | 80 | 0 | 0 | 40 |
| wSVM | Kernel: Sigmoid; gamma: 8; cost: 100 | No | 7 | 80 | 50 | 33 | 60 |
| wSVM | Kernel: polynomial; gamma: 0.25; cost: 100 | Yes | 4 | 0 | 0 | 67 | 20 |
| SVM | Kernel: Sigmoid; gamma: 0.25; cost: 0.001 | No | 7 | 100 | 0 | 0 | 50 |
| SVM | Kernel: polynomial; gamma: 0.25; cost: 0.001 | Yes | 4 | 100 | 0 | 0 | 50 |
| KNN | k neighbors: 23 | No | 7 | 100 | 0 | 0 | 50 |
| KNN | k neighbors: 23 | Yes | 4 | 100 | 0 | 0 | 50 |
| Random Forest | Number of trees: 2 | No | 7 | 40 | 100 | 67 | 60 |
| Random Forest | Number of trees: 2 | Yes | 4 | 60 | 0 | 33 | 40 |
PLS: Partial Least Squares algorithms (SIMFIT software v.6.9.9; www.simfit.org.uk); SVM: Support Vector Machines; KNM: K-Nearest Neighbors; GO/G1: Non-responders; G2: Partial responders; G3/G4: Responders. *we include the 3 or 6 best ranked by the prediction of the analyzed algorithms (N2, chr4q loss, chr15q11.1gain, chr15q14 loss, 17q21.31 gain, N1 and CEA). The best models found to predict the response to the RCT administrated prior to surgery are shown in bold.