A clinical prediction model integrating cardiovascular-kidney-metabolic biomarkers for the composite outcome of quality-of-life deterioration and rehospitalization in elderly HFpEF patients

Figure 1. Study flowchart and analytic pipeline. Patient selection flowchart: 1,847 screened → 302 enrolled in the development cohort (training 211 + internal validation 91), with CONSORT-style accounting of exclusions. A parallel temporal validation cohort (n = 102) assembled under identical criteria from consecutive admissions between January 2024 and June 2025 is shown alongside.

Figure 2. Restricted cubic spline analyses of six a priori continuous predictors in the training set. Adjusted hazard ratios with 95% confidence intervals (shaded bands) for the primary composite endpoint, modelled with four knots at the 5th, 35th, 65th, and 95th percentiles. Reference set at the median of each predictor. Models adjusted for age, sex, and NYHA class. (a) Uric acid: U-shaped relationship (P-nonlinear = 0.006), risk nadir near 350 μmol/L. (b) ln-NT-proBNP: J-shaped monotonically increasing curve (P-nonlinear = 0.042). (c) eGFR: inverse nonlinear association with plateau above 75 mL/min/1.73 m² (P-nonlinear = 0.035). (d) HbA1c: J-shaped relationship with risk escalating above 7.5% (P-nonlinear = 0.038). (e) TyG index: J-shaped curve with inflection near 9.0 (P-nonlinear = 0.11). (f) HGI: near-linear positive dose–response relationship (P-nonlinear = 0.38).

Figure 3. LASSO-Cox feature selection. (a) Coefficient profile plot showing the trajectory of each candidate variable’s coefficient as a function of log(λ); LVEF trajectory remains at zero across the entire λ path. (b) 10-fold cross-validation curve of partial likelihood deviance; vertical dashed lines mark λ.min (8 variables; primary model) and λ.1se (4 variables). (c) Selection frequency heatmap across the five multiply imputed datasets at λ.min; the eight predictors with 5/5 selection form the final model.

Figure 4. Nomogram and SHAP summary plot. (a) Cox regression nomogram constructed from the eight LASSO-selected predictors; each predictor axis is scaled to a common 0–100 “Points” scale at the top. The HbA1c axis is highlighted to indicate its role as the novel glycaemic predictor. The “Total Points” sum is mapped to 12-month and 24-month probabilities of the primary composite endpoint via the two bottom scales. (b) SHapley Additive exPlanations beeswarm summary plot. Each row represents a predictor sorted by descending mean absolute SHAP value; each dot represents an individual patient coloured by feature value (blue = low, red = high).

Figure 5. Model validation across development and temporal validation cohorts. (a) Time-dependent ROC curves at 12 months in the training (AUC 0.84) and internal validation (AUC 0.81) cohorts. (b) Time-dependent ROC curves at 24 months in the training (AUC 0.82) and internal validation (AUC 0.79) cohorts. (c) Calibration plot at 24 months in the training cohort (ICI = 0.024). (d) Calibration plot at 24 months in the internal validation cohort (ICI = 0.032). (e) ROC curves at 12 months in the temporal validation cohort (12-month AUC 0.80). (f) Calibration plot at 12 months in the temporal validation cohort (slope 0.87; ICI = 0.041).

Figure 6. Decision curve analysis across three pre-specified clinical decision scenarios. (a) Scenario 1 — intensified outpatient follow-up (threshold probability 15–25%); mean net benefit of the nomogram 0.142 at 20% threshold versus 0.089 for treat-all and 0 for treat-none. (b) Scenario 2 — initiation of CKM-directed pharmacotherapy (threshold 30–45%); mean net benefit 0.085 at 35% threshold. (c) Scenario 3 — multidisciplinary CKM referral (threshold 50–65%); mean net benefit 0.048 at 55% threshold. Treat all” assumes all patients will experience the outcome, and “Treat none” assumes none will. The threshold range (10-85%) represents the clinically plausible spectrum of risk where intervention is acceptable.

Figure 7. Within-stage stratification and subgroup performance. (a) Kaplan–Meier curves for the primary composite endpoint stratified by CKM stage 4a metabolic burden score (low 0–1, n = 88; intermediate 2–3, n = 138; high 4–6, n = 76). Log-rank χ² = 28.4; P < 0.001. 24-month cumulative incidences 21.6%, 41.3%, and 62.5%, respectively. (b) Forest plot of nomogram C-index across six pre-specified subgroups with interaction P-values. All P-interaction > 0.05.