Development and internal validation of prediction model for 90-day mortality in older patients with hip fracture using admission blood parameters

Figures

• 

  Fig. 1. Workflow for the development and validation of the clinical prediction model. The cohort of 1881 patients was split into a training set (n = 1317) for model development and an internal testing set (n = 564) for validation. The flowchart outlines the key stages: data preprocessing, feature selection, model construction, and a comprehensive performance evaluation assessing discrimination, calibration, and clinical utility. Abbreviations: AUC, area under the curve; DCA, decision curve analysis; LASSO, least absolute shrinkage and selection operator; MICE, multiple imputation by chained equations; NPV, negative predictive value; PPV, positive predictive value

• 

  Fig. 2. Comparative assessment of feature selection methodologies. A Feature importance ranking via the Boruta algorithm, categorizing features as confirmed important (green), tentative (yellow), unimportant (red), or shadow (blue). B, C Feature selection using LASSO regression. B Trajectory of variable coefficients across increasing regularization penalty (lambda). C Selection of the optimal lambda (dashed vertical line) via cross‑validation. D Venn diagram illustrating the overlap among features selected by stepwise regression, the Boruta algorithm, and LASSO. E Receiver operating characteristic (ROC) curves evaluating the discriminative performance of prediction models built from three feature subsets: the consensus of all three methods, the consensus of any two methods, and a hybrid set combining the two‑method consensus with the top five Boruta features. Abbreviations: CCI, Charlson Comorbidity Index; MCHC, mean corpuscular hemoglobin concentration; RDW-SD, red cell distribution width-standard deviation; RDW-CV, red cell distribution width-coefficient of variation; Lasso, least absolute shrinkage and selection operator; AUC, area under the curve

• 

  Fig. 3. Nomogram for predicting 90-day mortality in older patients with hip fracture. This nomogram is a graphical representation of the final logistic regression model, enabling the individualized estimation of 90-day mortality risk. For each predictor, a corresponding points value is assigned on the top scale. The sum of all points corresponds to a total points value, which is then projected onto the bottom scale to obtain the predicted probability of death. Abbreviations: CCI, Charlson Comorbidity Index

• 

  Fig. 4. Model performance and validation on training and internal testing sets. A-C Training set. A Receiver operating characteristic (ROC) curve; the area under the curve (AUC) quantifies discrimination. B Confusion matrix showing classification outcomes. C Calibration plot comparing predicted probabilities with observed event rates; the dashed line indicates perfect calibration. D-F Independent testing set. D ROC curve for external discrimination assessment. E Confusion matrix. F Calibration plot for probability accuracy validation

• 

  Fig. 5. Decision curve analysis for assessing the clinical utility of the prediction model. A Net benefit in the training set. B Net benefit in the independent testing set. In each panel, the net benefit of the model (blue or red solid line) is compared against the strategies of intervening for all patients (“Treat All”, gray dashed line) or for no patients (“Treat None”, black horizontal line). The model provides a positive net benefit within the threshold probability ranges where its curve lies above both reference lines. The consistent positive net benefit across both sets indicates the robustness of the model for supporting clinical decision-making