Serum growth differentiation factor 15 trajectory predicts 28-day mortality in critically ill patients: a multicenter cohort study

Ethical statement

This study was recorded in the Chinese Clinical Trial Registry (No. ChiCTR2400090770 and No. ChiCTR2400091418) and have been approved by the Institutional Study Board of the First Affiliated Hospital of Nanjing Medical University and the Second Affiliated Hospital of Harbin Medical University (2022-SR-678 and KY2024-066-01). The study was conducted in accordance with the principles of the Helsinki Declaration, and informed consent was obtained from all participants.

Study design and population

This study is a prospective, multicenter cohort conducted from March 2023 to August 2024 across three cities and four ICUs in China. Inclusion criteria included: 1. ≥18 years of age; 2. Enrollment within 24 h of ICU admission (timed from first arrival in the ICU); 3. At least one organ dysfunction, defined as a Sequential Organ Failure Assessment (SOFA) subscore ≥2 in any of the following systems: respiratory, cardiovascular, liver, renal, coagulation, or central nervous system; 4. An expected ICU stay of ≥48 h, as judged by the treating physician. Exclusion criteria included: (1) Expected death or withdrawal of life-sustaining treatment within 48 h after screening; (2) pregnant or breastfeeding; (3) hospitalized for >7 days prior to ICU admission. The study was structured into two phases: the development cohort phase, aiming to enroll 500 patients with a survival period of ≥7 days and serum samples available on days 1, 3, and 7. Following the development cohort, we continuously enrolled a validation cohort of 1,500 patients in two centers, requiring serum samples on at least day 1. The study followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline.

Clinical data and primary outcome

The research gathered the following patient information: (1) Baseline demographics. (2) Primary causes for ICU admission and clinical comorbidities. (3) Critical care scores: APACHE II and SOFA scores. (4) Laboratory indicators at ICU admission (D1), including markers of infection, indicators of organ function, and lactate. (5) Primary outcome: 28-day mortality rates.

Measurement of serum GDF15 levels

Venous blood specimens were collected using coagulation tubes and then left to rest at room temperature for 30 min. Samples were subjected to centrifugation at 4 °C and 3,000 rpm for 5 min within a 4-h window to isolate the serum. Serum samples from each center were collected and transported to the First Affiliated Hospital of Nanjing Medical University for unified measurement of GDF15 levels, utilizing a human GDF15 enzyme-linked immunosorbent assay kit (Cat# ELH-GDF15: RayBiotech, Guangzhou, China) following the protocol.

Calculation method for GDF15-load

To quantify the GDF15 levels over the first week of ICU stay, we introduced the concept of “GDF15-load” to measure the cumulative change. The formula for calculating GDF15-load is as follows.

$$GDF15-load={\displaystyle \frac{\left(GDF{15}_{D1}+GDF{15}_{D3}\right)}{2}\times \left(3-1\right)+{\displaystyle \frac{\left(GDF{15}_{D3}+GDF{15}_{D7}\right)}{2}\times \left(7-3\right).}}$$

Statistical analysis

For the development cohort, we employed GBTM to identify changes in serum GDF15 trajectories (Nagin, 2014; van der Nest et al., 2020). The GBTM was constructed by combining model fitting indicators with professional interpretability, which included proportions per class, Akaike information criterion (AIC), Bayesian information criterion (BIC), sample adjusted BIC (SABIC), and relative entropy (Ek).

Data of a continuous nature were depicted as average values ± the standard deviation (SD) when they followed a normal distribution, and as the median accompanied by the interquartile range (IQR) when they deviated from a normal distribution. For comparative analysis between groups, the Student’s t-test was employed for normally distributed data; otherwise, the Mann–Whitney U test or the Kruskal–Wallis test was used. Data of a nominal nature were reported in terms of counts and proportions (%) and examined through the application of the Chi-square test and Fisher’s exact test to determine statistical significance. For comparing subgroups, the analysis of covariance (ANOVA) and the Kruskal Wallis test were utilized, based on the general linear model. Spearman’s correlation test was used to analyze the correlation between GDF15 and other parameters.

Logistic regression analyses, including both univariate and multivariable models, were employed to evaluate the association between candidate predictors and 28-day mortality. For multivariable modeling, variables with statistical significance in univariate logistic regression—including demographics, disease severity scores, baseline laboratory indicators, and GDF15-D1—were included. To assess the independent association between GDF15 trajectories and 28-day mortality, we constructed three hierarchically adjusted multivariable logistic regression models: Model 1 adjusted for demographics; Model 2 further adjusted for disease severity; Model 3 additionally incorporated laboratory parameters reflecting key organ functions.

The predictive ability for clinical outcomes was evaluated by calculating the area under the receiver operating characteristic curve (AUROC), accompanied by 95% confidence intervals (CI). The Youden index was used to calculate the optimal cut-off value for predicting 28-day mortality. Kaplan-Meier curves were plotted to describe differences in clinical outcomes, with the log-rank test applied to assess the significance of differences. The values of Shapley Additive Explanations (SHAP), based on the XGBoost model, were applied to examine the importance of each predictor (Molnar, 2020). For missing values (less than 10%), a single interpolation method was used: mean imputation was employed for data exhibiting normal distribution, otherwise median imputation was utilized.

A p-value below 0.05 was deemed to indicate statistical significance, with all statistical tests being two-sided. The statistical evaluations were conducted utilizing R version 4.4.1 (R Core Team, 2024) and SPSS version 26.0 for analysis.

Clinical characteristics at baseline

A total of 1,973 patients were included in this study, with the enrollment and exclusion process detailed in Fig. 1. Baseline characteristics are detailed in Table S1. The median age was 61 years (IQR 51 to 71 years), with 63.1% (n = 1,244) being male. The median APACHE II score was 18 (IQR 15 to 21), with a mortality rate of 25.8% observed at 28 days. The baseline characteristics between the validation and development cohorts were largely comparable. However, the validation cohort exhibited lower alanine aminotransferase (ALT) and C-reactive protein (CRP) levels, alongside higher serum creatinine (SCr) and activated partial thromboplastin time (APTT) values.

GDF15 trajectories and derived subtypes in the development cohort

In the development cohort (n = 493), we used GBTM to fit the changes in serum GDF15 levels. The four-class trajectory model was selected for further analysis, demonstrating optimal performance based on AIC, BIC, and SABIC evaluations (Table S2).

Figure 2 illustrates the overall trajectories and the proportion of patients in each of the four subtypes. Based on the distinct patterns of GDF15 changes, we categorized the subtypes as follows: low-maintenance (LM), medium-maintenance (MM), high-increase (HI), and high-persistent (HP). The smallest LM subgroup had consistent GDF15 levels at approximately 2,000 pg/mL during the first week in the ICU. The MM group’s levels were steady at about 4,500 pg/mL. The largest HI subgroup saw a rise from 8,500 pg/mL at admission to 9,700 pg/mL by day 7. The smaller HP group started high at 21,000 pg/mL, dipping slightly to 20,700 pg/mL by day 3 before stabilizing. Detailed trajectories of GDF15 levels are provided in Table 1.

Further analysis of the clinical characteristics across the subtypes revealed significant differences in inflammatory markers, organ function, lactate, and critical care scores (Table 1). Patients in the HI and HP groups had higher levels of indicators of organ function, lactate, and critical care scores. Additionally, as initial GDF15 levels increased, the mean levels of aspartate aminotransferase (AST), total bilirubin (TBil), SCr, lactate, and critical care scores also increased, suggesting that patients with elevated GDF15 levels might experience poorer perfusion and organ function.

GDF15 subtypes and clinical outcomes

We evaluated 28-day mortality rates across GDF15 trajectory subtypes in the development cohort (n = 493) to gauge clinical outcome disparities. A marked increase in 28-day mortality rates was seen across GDF15 subtypes LM, MM, HI, and HP, correlating with ascending initial GDF15 levels (Table 1). Additionally, survival analysis indicated notable disparities in overall survival rates among the groups (Fig. 3A). However, further analysis revealed no significant differences in 28-day mortality and survival curves between the LM and MM subgroups, nor between the HI and HP subgroups (Table S3, Figs. 3B and 3C).

To determine whether these associations were independent of clinical confounders, we conducted multivariable logistic regression with hierarchical adjustments for demographics, disease severity, and laboratory variables (Table 2). After full adjustment, mortality differences remained significant, particularly between the HI/HP and LM subgroups, supporting the independent prognostic value of GDF15 trajectories.

GDF15-load reflects the cumulative change in GDF15 levels

Analyzing the features of trajectory subtypes, we observed that patients with sustained high GDF15 levels had worse outcomes at 28 days compared to those with lower or moderate levels, regardless of the trend. This suggests that the overall GDF15 level could be a key determinant of 28-day prognosis. We further quantified the cumulative change of GDF15 during the first week of ICU admission using GDF15-load. Significant differences in GDF15-load were observed across the four trajectory subtypes within the development cohort, showing a stepwise rise from the LM to the HP groups (Table 1). We integrated GDF15-load, critical care scores, and other parameters into a SHAP plot to evaluate their relative contributions to the predictive model. Results indicated that GDF15-load was the most significant predictor of 28-day outcomes (Fig. 4). In the receiver operating characteristic (ROC) analysis, predictive power of GDF15-load for 28-day outcomes yielded an AUROC of 0.703, not inferior to both APACHE II and SOFA scores (Fig. 5A). Survival analysis also indicated that individuals with elevated GDF15-load experienced notably reduced 28-day survival probabilities in the development cohort (Fig. S1).

These findings suggest that GDF15-load, derived from the GDF15 trajectory subtypes, serves as a powerful indicator for predicting 28-day outcomes in ICU patients.

Correlation between GDF15-D1 and GDF15-load

Trajectories or GDF15-load require continuous monitoring of GDF15. However, GDF15-D1, also identified as a significant contributor to prognosis in the SHAP plot, offers the advantage of being rapidly and easily available early on. We found a strong correlation of 0.778 between GDF15-D1 and GDF15-load using Spearman’s correlation coefficient in the development cohort (Fig. 5B). We then plotted the ROC curve to evaluate the predictive capacity of GDF15-D1 for 28-day outcomes, which yielded an AUROC of 0.692, still not inferior to both the APACHE II and SOFA scores (Fig. 5A). Based on the ROC curve of GDF15-D1, we calculated the cut-off value (7,582 pg/mL) to divide patients into high and low GDF15-D1 groups. The Sankey diagram showed strong agreement between the GDF15-D1-based groups and the trajectory-based groups (Fig. 5C).

We further validated the predictive ability of GDF15-D1 for 28-day mortality in the validation cohort (n = 1,480). Logistic regression revealed that GDF15-D1 serves as a substantial prognostic factor for 28-day outcomes (odds ratio (OR) = 1.115, 95% CI [1.089–1.141]) (Table S4). The AUROC for GDF15-D1 in predicting 28-day mortality was 0.723 (95% CI [0.692–0.754]), not inferior to the APACHE II score (AUROC = 0.717, 95% CI [0.687–0.747]) (Fig. S2). Survival analysis of the combined cohort (n = 1,973) demonstrated significantly reduced 28-day survival probabilities in individuals with elevated GDF15-D1 levels (Fig. S3).

These findings suggest that serum GDF15-D1 correlates well with GDF15-load and can reliably predict 28-day outcomes in ICU patients, making it a feasible and simplified alternative to the assessment of GDF15 trajectory and GDF15-load.

GDF15-D1 as a valid predictor of 28-day prognosis in postoperative ICU patients

Given the considerable heterogeneity of ICU populations, we then selected the postoperative subgroup with relatively good consistency, to further validate the predictive ability of GDF15-D1 in the validation cohort. This subgroup included 529 patients, with cardiovascular system (34.2%) and brain (24.4%) surgeries being the most common, collectively exhibiting an aggregate mortality rate of 18.3%. Baseline characteristics of these patients are detailed in Table S5. Survivors had a mean GDF15-D1 level of 5,386 pg/mL, contrasted with 12,952 pg/mL in non-survivors, signifying a significant increase in the latter group.

Logistic regression indicated that GDF15-D1 was a significant factor influencing 28-day outcomes in postoperative ICU patients, with an OR of 1.130 (95% CI [1.086–1.174]) (Table S6). The ROC curve for predicting 28-day prognosis also showed that GDF15-D1 had an AUROC of 0.734, showed no significant difference in AUROC between GDF15-D1 and APACHE II (AUROC = 0.728) or SOFA (AUROC = 0.711), indicating comparable predictive ability (Fig. 6A), marking a notable improvement over the heterogeneous ICU population. To further enhance prognostic accuracy, we constructed multivariable models incorporating GDF15-D1 with APACHE II or SOFA. The combined models significantly improved predictive performance (AUROC: GDF15-D1 + APACHE II, 0.816; GDF15-D1 + SOFA, 0.790), both statistically superior to each component alone (all p < 0.05; Fig. 6A). Notably, the integration of GDF15-D1 with both APACHE II and SOFA yielded the highest predictive performance (AUROC = 0.836), suggesting that GDF15-D1 may serve as a valuable complement to established severity scores in postoperative ICU patients. Survival analysis also revealed that individuals with elevated GDF15-D1 had notably reduced 28-day survival probabilities (Fig. 6B).

These findings suggest that serum GDF15-D1 is a strong predictor of 28-day outcomes in postoperative ICU patients, and its integration with existing severity scores can further enhance early risk stratification.