We conducted a retrospective observational study utilizing data from two independent cohorts. The derivation cohort was collected from the First Affiliated Hospital of Sun Yat-sen University database (FAH-SYSU) (Guangzhou, China), comprising data from 10,029 ICU admissions between January 2018 and December 2023. Access to clinical data was approved by the Clinical Research Ethics Committee of the First Affiliated Hospital of Sun Yat-sen University (Institutional Review Board number: 2022-048). Additionally, the validation cohort was obtained from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database, which contains a total of 299,712 ICU admissions at the Beth Israel Deaconess Medical Center (Boston, USA) between 2008 and 2019. One author (S. Li) is certified to get access to the database and was responsible for data extraction. As the data in the MIMIC-IV database were fully anonymized, the Institutional Review Board at the Beth Israel Deaconess Medical Center granted a waiver of informed consent (No.2001P001699).

Sepsis patients were diagnosed during ICU stay according to the third international consensus definition, that is, suspected infection accompanied by Sequential Organ Failure Assessment (SOFA) score 2 points higher than baseline (1). For patients with repetitive ICU admissions, only the first ICU admission was considered. Exclusion criteria included: age less than 18 years old, ICU length of stay less than 48 hours, less than 4 times of blood glucose measurements within the first 48 hours after ICU admission, and lack of follow-up information or unplanned discharge due to non-medical reasons. These criteria were applied to both cohorts.

The following variables were extracted: (1) demographic information, for instance, age, gender, surgical status, Charlson Comorbidity Index (CCI), disease severity scores (Acute Physiologic Assessment and Chronic Health Evaluation [APACHE] II score and SOFA score), (2) laboratory results and vital signs on ICU admission, including hemoglobin, white blood cell (WBC) count, platelet count, creatinine, total bilirubin (TBIL), prothrombin time (PT), and procalcitonin (PCT), (3) blood glucose levels within 48 hours of ICU admission, (4) 1-year mortality, 30-day mortality, ICU length of stay (ICU-LOS), the need of organ support (vasopressors, mechanical ventilation, continuous renal replacement therapy [CRRT]). All data were extracted using Structured Query Language (SQL) queries via the pgAdmin4 (version 6.15) interface for PostgreSQL.

The main outcome was all-cause 1-year mortality after ICU admission. Secondary outcomes included 30-day mortality, ICU mortality, ICU length of stay, vasopressor dosages, the administration of CRRT, ventilation duration, and doses of insulin.

The primary exposure in this study was blood glucose level trajectories within 48 hours after ICU admission. Specifically, we used the maximum blood glucose values in every 8-hour block, which were then standardized by subtracting the mean and dividing by the standard deviation, to form clusters with similar dynamic trends. In cases with missing glucose values, we imputed using the mean of each patient’s available measurements. Alternative methods (LOCF, NOCB, and MICE) showed consistent clustering and survival results, confirming the soundness of this approach.

Next, we utilized K-means clustering to group patients into distinct clusters, employing the Scikit-learn package (version 1.5.2). This study was conducted according to the CAIR checklist ( Appendix 1 ) (22).To determine the optimal number of clusters (K), multiple metrics were examined, including the sum of squared errors (SSE), Silhouette Score, Calinski-Harabasz index, and Davies-Bouldin index (23–25). Besides, an elbow plot was generated to visualize the K and SSE, assisting in identifying a point where increasing K no longer significantly improved the fit. The final choice of K was based on a consensus among these measures and a qualitative inspection of the elbow plot. Each patient was then assigned to the resulting clusters, and subsequent analyses were conducted to compare clinical characteristics and outcomes across these groups.

Moreover, we performed an external validation using the MIMIC-IV database, ensuring the generalizability of the derived clusters. The same data preprocessing and clustering processes were replicated in this cohort.

To improve transparency and reproducibility, we incorporated a schematic workflow diagram ( Figure 1 ) to visually summarize the study design and analytical pipeline. This diagram illustrates the key stages of the study: cohort selection and exclusion criteria, preprocessing of dynamic BG measurements, trajectory clustering using k-means, external validation in an independent cohort, and the subsequent outcome analyses including 1-year mortality comparisons across clusters. This visual representation helps clarify the sequence and rationale of the analytical steps.

Continuous variables were reported as mean (standard deviation [SD]), while categorical variables were reported as counts with percentages. Continuous variables were compared with a one-way ANOVA test or the Kruskal-Wallis test based on the normality of candidate variables. Categorical variables were compared with either a chi-squared test or Fisher’s exact test when appropriate. Kaplan-Meier curves and log-rank tests were used to compare the 1-year survival outcomes across the groups. Additionally, we used both univariate and two multivariate Cox proportional hazards regression models (Model 1: adjusted for age and gender; Model 2: adjusted for age, gender, admission department, CCI, APACHE II, SOFA, creatinine, total bilirubin, hemoglobin, heart rate, mean arterial pressure, and urine output) to inspect the association between the formed clusters and 1-year survival risks.

To further validate the credibility of the identified clusters, subgroup analyses were performed stratifying by age (<=65 and > 65 years), gender, presence of septic shock, and surgical status, in order to evaluate the hazard ratios of 1-year morality within each subgroup. We also conducted a sensitivity analysis among patients with and without pre-existing diabetes to assess whether the clusters were associated with similar outcomes regardless of diabetes status. The significance of the interaction was assessed using the Wald test, and P values for interaction were reported. A P less than 0.05 was considered indicative of a statistically significant interaction. Lastly, restricted cubic spline (RCS) with four knots was adopted to visualize the potential non-linear relationships between two glycemic indexes (mean blood glucose levels, and the coefficients of variation [CV] within the 48 hours after ICU admission) and 1-year mortality. CV is an indicator for glucose variability in continuous glucose monitoring (26), calculated as the standard deviation of glucose values divided by the mean values in the same observation period.

Statistical significance was defined as a two-tailed P value of less than 0.05. All statistical analyses were conducted using R (version 4.2.2) or Python (version 3.12.2).

A total of 4,608 patients diagnosed with sepsis were admitted to FAH-SYSU between January 2018 to December 2023. After screening, 953 patients were excluded, with 3,655 eligible patients included in our study. The process of participant enrollment is displayed in Supplementary Figure S1A . Among these patients, the mean age was 58.71 years (SD: 16.20), with 70.26% as male patients. The mean APACHE II score at admission was 20.39 (SD: 7.76), and 54.31% presented with septic shock (N=1,985), indicating high disease severity. The mean blood glucose level at admission was 164.82 mg/dL (SD: 75.01). There were 1,674 (45.80%) patients who died within 1 year after diagnosis with sepsis, as shown in Table 1 .

To determine the optimal number of clusters (K), multiple metrics were employed, as shown in Supplementary Table S1 and Supplementary Figure S2 . The elbow plot and SSE values suggested an “elbow” around K=5, where the incremental decrease of SSE became relatively small after this point ( Supplementary Figure S2 ). Additionally, the Calinski-Harabasz index peaked at K=5, while the Silhouette Score remained satisfactory. Consequently, we selected K=5 as the optimal cluster number, as these distinct trends exhibited high internal similarity, with additional clusters offering limited gain in differentiation.

Figure 2A illustrates the longitudinal BG trajectories of the five identified clusters over the initial 48 hours following ICU admission. Cluster 0 (30.64%) displayed a relatively stable and low glucose trajectory over the entire course (“low-stable”), with a median value of 127.8 mg/dL (IQR: 108.0–149.0) at baseline and 120.6 mg/dL (IQR: 104.4–136.8) at 48 hours. Cluster 1 (9.99%) maintained high glucose levels throughout the period (“high-stable”), with a median decrease from 288.0 mg/dL (IQR: 241.2–347.4) to 255.6 mg/dL (IQR: 223.2–302.4), corresponding to an absolute reduction of about 32.4 mg/dL. Cluster 2 (31.66%) started at a moderate glucose level and remained stable (“moderate-stable”), showing only a slight decline from 158.4 mg/dL (IQR: 135.0–183.6) to 156.6 mg/dL (IQR: 138.6–178.2). In contrast, Cluster 3 (10.75%) presented with a high initial glucose level but showed a pronounced drop over the 48-hour period (“high-decreasing”), with values falling from 271.8 mg/dL (IQR: 239.4–320.4) to 167.4 mg/dL (IQR: 140.4–198.0), representing an absolute decrease of approximately 104.4 mg/dL and about 39% reduction from baseline. Lastly, cluster 4 (16.96%) was characterized by starting at a moderate level and then increasing (“moderate-increasing”), which showed an upward trend in BG, increasing from a median of 183.6 mg/dL (IQR: 153.0–214.2) to 205.2 mg/dL (IQR: 180.0–237.6), with an overall increase of 21.6 mg/dL.

The baseline characteristics of these clusters are summarized in Table 1 , revealing diverse demographics, vital signs, blood counts and biochemistry profiles. Notably, cluster 1 (“high-stable”) had the highest mean glucose levels over the period (268.06 mg/dL), whereas glycemic variability in cluster 3 (“high-decreasing”) was the greatest (CV: 0.28), further supporting the clinical relevance of our clustering approach. Additionally, cluster 1 exhibited the highest mean APACHE-II score (23.21) and 1-year mortality (55.89%), followed by cluster 4 (“moderate-increasing”) and cluster 3 (52.58%, 51.65%, respectively). Cluster 0 (“low-stable”) was related to the lowest 1-year mortality (39.11%). Kaplan-Meier curves of all clusters are illustrated in Figure 2B , indicating that cluster 1 showed the worst 1-year survival (log-rank P < 0.001).

Primary and secondary outcomes across these clusters are displayed in Table 2 . There was no significant difference between these trajectories and ICU length of stay (P = 0.165). The trends of ICU and 30-day mortality across the subtypes were similar to 1-year results. Specifically, individuals in cluster 3 were associated with the largest amount of vasopressor usage (50 mg, IQR [10, 148]), along with the longest ventilation durations (106 hours, IQR [51, 212]), suggesting a greater need for organ support and hemodynamic stabilization for those with a “high-decreasing” glucose trajectory. Additionally, cluster 1 (“high-stable”) received the highest doses of insulin (320 IU, IQR [168, 580]).

Univariate analysis revealed significant risk factors associated with 1-year mortality in sepsis patients, including age, gender, admission department, CCI, APACHE II, SOFA, creatinine, total bilirubin, hemoglobin, heart rate, mean arterial pressure, and urine output. The association between blood glucose trajectories and 1-year mortality risks in different Cox proportional hazards regression models is presented in Table 3 . Both unadjusted and adjusted hazard ratios demonstrated that cluster 1 (“high-stable”) was related to a significantly higher risk of 1-year mortality compared to cluster 0 (“low-stable”) (HR = 1.68, 1.63, 1.61 in the univariate model, model 1, and model 2, respectively; all P < 0.001). Similarly, both cluster 3 (“high-decreasing”) and cluster 4 (“moderate-increasing”) were linked to increased 1-year mortality risks in the full multivariate model (HR: 1.38, 95% CI: 1.16-1,65, P < 0.001; HR: 1.37, 95% CI: 1.18-1.60, P < 0.001, respectively). No significant difference was found in the 1-year mortality risk of cluster 2 (“moderate-stable”) compared to the reference group (P = 0.077 in model 2).

To externally validate our clustering approach, we applied the same process to 10,874 sepsis patients from the MIMIC-IV database, with the selection process outlined in Supplementary Figure S1B . Baseline characteristics are summarized in Supplementary Table S2 . Consistent clustering subtypes were obtained in the validation cohort: cluster 0 (“low-stable”, 53.27%), cluster 1 (“high-stable”, 4.71%), cluster 2 (“moderate-stable”, 25.51%), cluster 3 (“high-decreasing”, 10.27%), cluster 4 (“moderate-increasing”, 6.24%). As displayed in Figure 3 , cluster 0 exhibited the most favorable 1-year survival, whereas both cluster 1 and cluster 4 showed the worst 1-year survival outcomes (log-rank P<0.001).

To control baseline imbalances across the formed clusters, subgroup analyses were conducted in the derivation cohort, as illustrated in Figure 4 . The association between blood glucose trajectories and 1-year mortality risk remained largely consistent regardless of gender and surgical status, where a “low-stable” or “moderate-stable” glucose trend was correlated with better outcomes. In aging patients (>65 years), only cluster 1 (“high-stable”) and cluster 3 (“high-decreasing”) showed significant long-term mortality risks (P=0.006, 0.002, respectively), suggesting that high glucose levels at admission may indicate poor outcomes for elderly individuals. Additionally, in septic shock patients, only cluster 1 exhibited a significantly worse prognosis (P=0.022), which could be related to increased oxidative stress, coagulation activation, and endothelial dysfunction as a result of persistent stress hyperglycemia.

To validate the impact of diabetes history on the BG clustering process, sensitivity analysis was performed in diabetes and non-diabetes patients, respectively. In patients without diabetes history (N=2,981), the survival outcomes across trajectories were similar to the main results ( Supplementary Figures S3A, B ). However, in diabetic patients, the blood glucose trajectories were similar to the primary analysis, whereas 5 clusters showed different 1-year mortality. Cluster 0 (“low-stable”) presented the worst 1-year prognosis, though there was no statistically significant difference in 1-year mortality rate among the five groups (log-rank P = 0.2) ( Supplementary Figure S3C, D ). This suggests that diabetic patients may have a distinct immune response and metabolism as a consequence of the host’s chronic tolerance to hyperglycemia.

Lastly, RCS was used to further explore the optimal mean blood glucose and CV ranges associated with lower 1-year mortality risks ( Figure 5 ). Both blood glucose and CV exhibited a nonlinear relationship with 1-year mortality (P for nonlinear < 0.001, both). There was a U-shaped curve between the mean blood glucose values within 48 hours and 1-year mortality, with an optimal range between 122 mg/dL and 160 mg/dL. Contrarily, the relationship between CV and 1-year mortality exhibited a J-shaped curve, with a higher risk of death when CV was greater than 0.18.