Skip to main content

Table 1 Summary of strategies to build prognostic survival models

From: A strategy to build and validate a prognostic biomarker model based on RT-qPCR gene expression and clinical covariates

Strategy Variable selection method Functional relationship Survival model Coefficients shrinkage
Uni_Cox-[1-9]* Each candidate covariate was selected apriori and included in a univariate model.There is 1 suffix number in the model name per selected covariate All candidate covariate are continuous, and a linear relationship is assumed Cox model  
bwAIC_FP_Cox-[1-5]* bwAIC_FP/C2-3Fac_Cox-[1-5]* Backward elimination using AIC criterion Fractional polynomial to model functionalrelationship. The suffix in the model namecorresponds to the degree of flexibility,controlled with the parameter α=0.05, 0.1,0.2,0.3,0.4. In the bwAIC_FP/C2-3Fac_Cox-[1-5] models, variable C2 and C3 are dichotomized according to expert knowledge. Larger a values correspond to more flexible relationships. Cox model  
MFP_Cox-[1-15]* MFP procedure for variable selection controlled by the parameter select =0.05,0.10, 0.15. Larger select values correspond to less stringent variable selection. Suffix in the model name corresponds to a combination of select and alpha values Fractional polynomial to model functional relationship. The parameter α controls the degree of flexibility, α=0.05,0.1,0.2,0.3,0.4. Larger a values correspond to more flexible relationships Cox model  
Lasso-[1-5]* Lasso_C2-3Fac-[1-5] L1 penalty for variable selection, controlled by the parameter λ=0.01,0.1,1,10,100. Larger λ values correspond to sparser models. Suffix in the model name correspond to the level of penalization. Linear relationship, except for C2 and C3 which have been dichotomized according to expert knowledge in the Lasso_C2-3Fac_Cox- models. Lasso Cox model L1 penalty
aLasso-[1-5]* aLasso_C2-3Fac-[1-5]* L1 penalty for variable selection, controlled by the parameter λ=0.01,0.1,1,10,100. Larger λ values correspond to sparser models. Suffix in the model name correspond to the level of penalization. Linear relationship, except for C2 and C3 which have been dichotomized according to expert knowledge in the aLasso_C2-3Fac_Cox- models. Adaptive LassoCox model Adaptive lasso penalty for coefficients shrinkage (larger coefficients are less shrinked towards 0 in the adaptive lasso model than in the lasso model.
SCAD-[1-5]* SCAD_C2-3Fac-[1-5]* L1 penalty for variable selection, controlled by the parameter λ=0.01,0.1,1,10,100. Larger λ values correspond to sparser models. Suffix in the model name correspond to the level of penalization. Linear relationship, except for C2 and C3 which have been dichotomized according to expert knowledge in the SCAD_Lasso_C2-3Fac_Cox- models. SCAD Cox model SCAD penalty for coefficients shrinkage (larger coefficients are less shrinked towards 0 in the SCAD model than in the lasso model.
Lasso_Cox-[1-5]* Lasso_C2-3Fac_Cox-[1-5]* L1 penalty for variable selection, controlled by the parameter λ=0.01,0.1,1,10,100. Larger λ values correspond to sparser models. Suffix in the model name correspond to the level of penalization. Linear relationship, except for C2 and C3 which have been dichotomized according to expert knowledge in the Lasso_C2-3Fac_Cox- models. Cox model  
  1. The first column gives the names of the tested strategies. The strategies cover a wide range of state-of-the-art methods from both low and high dimensional settings. The second column details the variable selection method used; the third column the functional relationship for continuous covariates; the fourth column the survival model; and the last column the coefficients shrinkage strategy if any. The suffix indicates the index of the prognostic model falling in the strategy. For example, Uni_Cox −[1−9] means that 9 univariate Cox models were built, each of them being suffixed by digit 1 to 9.