Table 3.
Performance metrics of different Cox regression models where proteins were selected for inclusion by using a two-step random-split approach in the MDCS-CVA (N = 3,918).
| Model | C-statistic | Change in C-statistic from clinical model | P-value (LR) | Change in C-statistic from clinical + strongest marker model | P-value (LR) |
|---|---|---|---|---|---|
| Clinical model* | 0.7222 | – | – | ||
| Clinical + strongest marker** | 0.7284 | 0.0062 | < 0.0001 | – | – |
| Clinical + all proteins*** | 0.7705 | 0.0483 | < 0.0001 | 0.0421 | < 0.0001 |
| Clinical + Coxa | 0.7379 | 0.0157 | < 0.0001 | 0.0095 | < 0.0001 |
| Clinical + StepwiseCoxb | 0.7373 | 0.0151 | < 0.0001 | 0.0089 | < 0.0001 |
| Clinical + LassoCoxc | 0.7492 | 0.0270 | < 0.0001 | 0.0208 | < 0.0001 |
| Clinical + RSFd | 0.7436 | 0.0241 | < 0.0001 | 0.0152 | < 0.0001 |
*Covariates included in the model were age, sex, smoking status, BMI, educational level, history of hypertension, prevalent diabetes mellitus, C-reactive protein, HbA1c, and LDL-cholesterol.
**Covariates and growth/differentiation factor-15 (GDF-15).
***Covariates and all proteins (n = 138).
aCovariates and 9 proteins (AR, CXCL9, GDF15, GH, HE4, IL6, NTproBNP, SCF, UPAR) associated (P < 0.05) with all-cause mortality in a Cox regression model after adjustment for covariates in both random samples of the MDCS-CVA.
bCovariates and 6 proteins (GDF15, CASP3, EGFR, EZR, HE4, NTproBNP) associated (P < 0.05) with all-cause mortality with mutual adjustment in both random samples of the MDCS-CVA using a stepwise Cox regression with backwards elimination of proteins with P < 0.05. MPO was excluded due to diverging associations with all-cause mortality in the two random samples.
cClinical variables and 13 proteins (CXCL9, EGFR, EZR, GDF15, GH, HE4, KLK6, MB, NTproBNP, SCF, TIM, TRAIL, UPAR) retained in both random samples of the MDCS-CVA using a Lasso penalized Cox regression and lambda minimum for protein selection.
d Clinical variables and 21 proteins (FABP4, FasL, GDF15, HE4, HGF, IL12, IL6, mAmP, MMP1, MMP12, MYD88, NTproBNP, PRSS8, PSGL1, PTPN22, PTX3, RAGE, REN, SCF, THPO, TIM) retained in both random samples using a RSF backward elimination approach.