Fig. 2.

Predictive score of mortality post hip-fracture.

(a) Variables measurable prior to surgery capable of predicting death were identified using a lasso-regularized logistic regression model with 10-fold cross validation. To render the analysis more robust we repeated the model fitting (n = 500). Bar diagram depicts the frequency of models identifying a given variable. Only variables being identified in 2/3 of the models were selected for further analysis (dashed line).

(b) Scatter plot of a two components of a Partial Least Square Discriminant Analysis (PLS-DA) identifying the degree of variance of the predicted variable (death) explained by the two predictive variables identified by regularized logistic regression (neopterin concentration in nmol/L and % CD16⁺ CD56dimNKG2C⁺ NK cells). The explained variance is 41.2% and 0.2% for component 1 and 2 (PLS-C1 & PLS-C2), respectively. Symbols discriminate between alive (white symbols) and dead (black symbols) hip-fracture patients one-year post-surgery.

(c) Scatter plot depicting a mortality score mathematically derived from the PLS component 1 stratified according to survival status. The mortality score is calculated as the sum of neopterin concentration (nmol/L) and the % CD16⁺ CD56dimNKG2C⁺ NK cells. A proposed cut-off value (24.7) is indicated with a dotted line; this gives a test with 100% sensitivity and 70% specificity. Statistical significance is calculated with a non-parametric Mann-Whitney test.

(d) Receiver operating characteristic (ROC) curves of logistic regression models of death occurrence predicted by the two predictive variables in combination or individually (blue line for % CD16⁺ CD56dimNKG2C⁺ NK cells at D0; redline for plasmatic concentration of neopterin (nmol/L) at D0 and black line for the 2 variables in combination). Area under curve (AUC) for each curve is indicated.