Dynamic association between triglyceride-glucose-body mass index and hospital mortality in critically ill patients with heart failure: a multicenter retrospective cohort study

Fig. 1. Flow diagram of patient selection from three critical care databases. The diagram illustrates the sequential application of inclusion and exclusion criteria, including identification of heart failure diagnoses, age restriction, selection of first admissions, ICU length of stay requirement, and completeness of TYG-BMI data, with final sample sizes for MIMIC-III, MIMIC-IV, and eICU databases

Fig. 2. Temporal changes in TYG-BMI across three databases during the first 7 days of ICU admission. Data are presented as mean values with 95% confidence intervals. Each line represents a different database (eICU, MIMIC-III, and MIMIC-IV), demonstrating the dynamic trajectory of TYG-BMI following ICU admission in patients with heart failure

Fig. 3. Restricted cubic spline analysis of the association between Day 0 TYG-BMI and hospital mortality. Panels A–C show results from Model 1 (adjusted for age and gender) in MIMIC-IV, MIMIC-III, and eICU, respectively; panels D–F show corresponding results from Model 2 (fully adjusted model, additionally adjusted for SOFA score, Charlson comorbidity index, insulin use, vasopressor use, mechanical ventilation, albumin, creatinine, and white blood cell count). The red solid line represents the odds ratio point estimate, blue dashed lines represent the 95% confidence interval, and the gray horizontal dashed line denotes the OR = 1 reference. The reference value was set at the median TYG-BMI of each database, and the optimal number of knots was selected from 3 to 7 based on the Akaike information criterion. Rug plots at the bottom indicate the distribution density of the sample

Fig. 4. Three-dimensional surface plots and heatmaps depicting the relationship between TYG-BMI intensity, time since ICU admission, and hospital mortality. Panels show three-dimensional surface plots of predicted mortality probability from A MIMIC-IV, C MIMIC-III, and E eICU databases, with warmer colors indicating higher mortality risk. Corresponding odds ratio deviation heatmaps are presented for B MIMIC-IV, D MIMIC-III, and F eICU databases, where blue indicates protective effects (OR < 1) and red indicates increased risk (OR > 1), with dashed contour lines denoting OR = 1. Surfaces and heatmaps were generated using generalized additive models with tensor product smooth functions for the TYG-BMI × time interaction

Fig. 5. Forest plot of odds ratios for hospital mortality by TYG-BMI range across three databases. Results are stratified by days since ICU admission (Day 0 through Day 7) and presented with 95% confidence intervals on a logarithmic scale. Different colors and shapes represent distinct databases, facilitating comparison of effect estimates across cohorts

Fig. 6. Sensitivity analyses of the TYG-BMI–time–mortality relationship using three-dimensional surface plots. Panels A–C show the TYG-BMI–time–mortality surfaces from MIMIC-IV, MIMIC-III, and eICU databases, respectively, under the missing rate < 30% sensitivity analysis. Panels D–F show the corresponding surfaces restricted to the first 3 days after ICU admission (Day 0–3) for MIMIC-IV, MIMIC-III, and eICU databases, respectively. All surfaces were generated using generalized additive models with tensor product smooth functions, with warmer colors indicating higher predicted mortality