This study was performed in strict accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement (15). Our study was approved by the Independent Ethics Committee for Clinical Research and Animal Trials of The First Affiliated Hospital, Sun Yat-sen University (approval No.: [2022]700). Informed consent was waived as this study did not involve any possible damage to the enrolled patients. The study was registered in China Clinical Trial Registry (Registration No.: ChiCTR2300072040).

A single-center, retrospective, observational study was conducted in a general and surgical ICU with 56 beds in a tertiary hospital from June 2016 to May 2020. Only the first ICU stay was considered if a patient had multiple ICU admissions. Patients with absence of case data, age less than 18 years, acute cardiovascular events during hospitalization (successful resuscitation after respiratory and cardiac arrest, acute coronary syndrome, aortic dissection, pulmonary embolism), cardiac and great vessel surgery, and rhabdomyolysis were excluded. Only adult patients with myoglobin measurement within 48 h after entering the ICU were included to describe the distribution of myoglobin levels, and patients with at least three myoglobin measurements within 7 days after ICU admission were included for trajectory.

Demographic characteristics, clinical parameters, and laboratory examination results were extracted from the electronic databases of hospital information system and ICU system. The maximum myoglobin measurement before entering the ICU was taken as baseline value, and only the maximum value was retained for multiple measurements within the same day (13). Other laboratory indicators, including creatine kinase MB (CK-Mb), pro-B-type natriuretic peptide (pro-BNP), and cardiac troponin I (cTnI), were also collected. In addition, the maximum values of lactate, total bilirubin, and creatinine, the lowest urine volume, and mean arterial pressure (MAP) were recorded 72 h after entering the ICU to evaluate the tissue perfusion and organ functional status of the patients.

In-hospital death was determined based on the information in the electronic database (extracted from the home page of the medical record). Sepsis and Sequential Organ Failure Assessment (SOFA) scores were defined according to sepsis 3.0 (16). During assessment, the Glasgow Coma Scale (GCS) score was found to be subjective and heterogeneous; hence, it was removed from the SOFA score. Acute kidney injury (AKI) was diagnosed in line with the Kidney Disease Improving Global Outcomes (KDIGO) criteria (17). ACS (ST-segment elevation myocardial infarction, non-ST-segment elevation myocardial infarction, and unstable angina pectoris) was diagnosed by cardiovascular physicians based on electrocardiographic findings, troponin level, and previous cardiac conditions. Other major diagnoses were made based on the codes of the International Classification of Diseases (ICD), Ninth Edition. Other ICU treatment records were derived from the electronic database (extracted from the ICU system).

Two practitioners were responsible for assessing the integrity and quality of data of the included patients. The data integrity was evaluated mainly for the in-hospital survival. The quality assessment included the diagnostic accuracy of sepsis and AKI to control for the disease misclassification. Simultaneously, sampling quality control was carried out by two senior physicians.

Based on myoglobin levels (ng/mL) within 48 h after ICU admission, the patients were divided into four groups, namely normal (<150), mild increase (150–500), middle increase (500–1,000), and severe increase (≥1,000), consistent with the cutoff values previously utilized by other authors (18, 19). Differences between the groups were analyzed. The primary outcome was evaluated with comparison of in-hospital mortality among each Trajectory group, and the secondary outcome was 28-day mortality. Percentages were used to describe discrete variables, and chi-squared or Fisher’s exact test was applied to test hypothesis. Medians and interquartile ranges (IQRs) were used to describe continuous variables, and Kruskal–Wallis rank-sum test was applied to test hypothesis among the four groups. In LCTM analysis, the population was classified based on daily myoglobin examinations from before ICU admission to 7 days after ICU admission. Since there were very few maxima of myoglobin concentration, the examination values were transformed logarithmically. The optimal number of latent classes was determined by relative entropy, Bayesian Information Criterion (BIC), and the proportion of each latent class. The percentages of subjects classified with a posterior probability above 0.7/0.8/0.9 and the table of posterior classification were also demonstrated. Based on BIC and relative entropy, we finally selected the number of 4. The 28-day Kaplan–Meier (KM) curve was created, and log-rank test was applied to test the hypothesis. To verify the robustness of the above model in different populations, subpopulations with or without sepsis, and with or without surgery were filtered to re-fit the LCTM algorithm, and the corresponding KM curves were created. In sensitivity analysis, patients with survival time less than or equal to 72 h were excluded. The consistency between the sensitivity analysis and the original analysis was calculated. To intuitively demonstrate the effect of covariates on the patients’ survival time, restricted mean survival time (RMST) regression based on pseudo values was used to assess the differences in 28-day RMST (20). Age over 65 years, sex, SOFA score, sepsis, chronic kidney disease (CKD), hepatic disease, post-surgery, and minimum MAP within 72 h were incorporated into the regression as confounders. Forest plot was also created. Random-forest missing-value imputation method was applied to impute missing values (21). In this study, missing value imputation was only applied in regression analysis. Detailed description of LCTM and RMST regression is shown in Supplementary Statistical Methods. All statistical analyses were performed with R 4.2.1, and two-sided p < 0.05 was considered statistically significant. Bonferroni’s correction was applied for multiple comparisons.

A total of 6,872 patients were included to describe the distribution of myoglobin concentrations (Figure 1). The median age of the patients was 54 years, 58.6% were male, the overall in-hospital mortality rate was 10.1%, and the median SOFA score was 7 points. Of the enrolled patients, 56.5% (n = 3,886) had elevated myoglobin concentration when entering the ICU, including 31.5% with mild increase, 12.0% with middle increase, and 13.0% with severe increase (Figure 2A). There was no difference in the proportion of patients with past medical history of chronic cardiovascular diseases among the groups. As shown in Table 1, the severe increasing group showed significantly higher proportion of sepsis (19.7% vs. 8.7%, p < 0.001), in-hospital mortality (24.2% vs. 6.6%, p < 0.001), and use of vasoactive drugs (31.9% vs. 64.3%, p < 0.001) than the normal group. Only the patients in the severe increasing group had a higher 28-day mortality than those in the normal group (Figure 2B).

A total of 2,448 patients were enrolled in myoglobin trajectory modeling (Figure 1). The median age of the patients was 59.0 years, 66.7% were male, the overall in-hospital mortality rate was 17.8%, and the median SOFA score was 8.0.

Considering the BIC and relative entropy of the trajectory model, four groups were finally determined (Supplementary Tables 1–3). The four trajectory groups were as follows: Group 1 (65.6%, n = 1,606), slight fluctuation of myoglobin level within the normal range (hereinafter referred to as the steady group); Group 2 (21.4%, n = 523), initially high myoglobin level rapidly decreased to normal (the gradually decreasing group); Group 3 (11.1%, n = 272), myoglobin level slightly increased and then decreased slowly (the slowly rising group); and Group 4 (1.9%, n = 47), myoglobin level rapidly increased and then decreased gradually (the rapidly rising group). All dynamic trajectories are shown in Figure 3A. In the sensitivity analysis, the individuals whose survival time was less than or equal to 72 h were excluded, and a new LCTM was fitted (Supplementary Figure 1). The new model classified 98.2% of the individuals in the same way as the original model. In the new model, the proportion of patients in the same group was 92.7% in the rapidly rising group and 95.9% in the slowly rising group.

The characteristics and outcomes of the patients based on LCTM are shown in Table 2. In terms of demographic characteristics, the median age in the rapidly rising group and the steady group was higher than that in the gradually decreasing group and the slowly rising group. There was a significant difference in gender distribution among the four groups (p < 0.001), with lower proportion of male patients in the steady group than in the other three groups. The proportions of chronic cardiovascular diseases, diabetes, and CKD were lower in the rapidly rising group than in the steady group, but the rapidly rising group had the highest SOFA score, lactic acid, and bilirubin, the lowest MAP, and the highest incidence of AKI and sepsis. In terms of outcomes, there was a significant difference in 28-day mortality among the groups (p < 0.001). In the gradually decreasing group, myoglobin level was the highest at admission to the ICU, and the in-hospital mortality was the lowest among all groups. The in-hospital mortality in the slowly rising group and rapidly rising group was higher than that in the steady group. There was no significant difference in the in-hospital mortality between the steady group and the gradually decreasing group.

To evaluate the stability of the model in different subgroups, we further studied model establishment in sepsis/non-sepsis (n = 587/1,861) and surgery/non-surgery (n = 1,611/837) patients based on previous investigation (22–24). The trajectory plots and the corresponding KM curves are shown in Supplementary Figures 2, 3. In the above four subgroups, the myoglobin trajectories and the trend of the KM curves were both similar to the overall pattern.

With the steady group as the reference, RMST regression based on pseudo values was employed to evaluate the relative changes in the expected 28-day survival time between different trajectory groups. Multivariate regression results are shown in Figure 4. After adjustment for gender, age, and other confounders, there was no statistically significant difference in the expected 28-day RMST between the gradually decreasing group and the steady group. However, compared with the expected 28-day RMST of the steady group, that of the slowly rising group and the rapidly rising group decreased by 4.25 days and 12.08 days, respectively (p < 0.001).

Of 2,448 patients, 12.7% had missing values, which were deleted in the RMST regression. The regression results after filling the missing values by a random-forest algorithm were similar to those without imputation (Supplementary Figure 4).