Longitudinal analysis of ARDS variability and biomarker predictive power in burn patients Authors: Jaechul Yoon, Dohern Kym, Yong Suk Cho, Jun Hur, Dogeon Yoon Published: 2024 ABSTRACT AbstractAcute Respiratory Distress Syndrome (ARDS) is a severe complication in burn patients, characterized by rapid lung inflammation and hypoxemia. Managing ARDS is challenging due to its high mortality rates and the complex interplay of various pathophysiological factors. This study aims to investigate the heterogeneity of ARDS in burn patients admitted to the Intensive Care Unit (ICU) and evaluate the predictive efficacy of biomarkers in comparison to the PaO2/FiO2 (PF) ratio. A retrospective cohort study was conducted at the Burn Intensive Care Unit of Hangang Sacred Heart Hospital in Seoul, Korea, from July 2010 to December 2022, involving 2,318 patients. Longitudinal k-means clustering was employed to identify patient subgroups based on PF ratio trajectories. Biomarkers, including albumin, white blood cell (WBC) count, and lactate, were assessed for their predictive accuracy using multivariable logistic regression, area under the curve (AUC) analysis, net reclassification improvement (NRI), and integrated discrimination improvement (IDI). Four distinct patient clusters were identified. Cluster A exhibited the highest mortality rate at 54.9%, while Cluster D showed the lowest at 7.9%. Albumin demonstrated significantly higher predictive accuracy in Cluster A (AUC: 0.905, NRI: 0.421) and Cluster C (AUC: 0.915, NRI: 0.845), outperforming the PF ratio in both groups. However, in Clusters B and D, biomarkers did not significantly improve upon the predictive power of the PF ratio. Across all clusters, the integration of biomarkers with the PF ratio led to modest improvements but did not consistently outperform the PF ratio as a standalone predictor. This study reveals substantial heterogeneity in ARDS progression among burn patients, with varying mortality rates and biomarker efficacy across clusters. While biomarkers such as albumin showed potential in specific subgroups, their overall contribution to predictive accuracy was limited. Further multicenter, prospective studies are required to validate these findings and develop more refined predictive models. Personalized treatment strategies, based on biomarker profiles and traditional clinical metrics, could enhance ARDS management in burn patients. Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-024-77188-x.Keywords: ARDS, Burn patients, Heterogeneity, Biomarkers, Precision medicine, Retrospective cohort study, Clustering analysisSubject terms: Medical research, Risk factors Background Acute Respiratory Distress Syndrome (ARDS) is a critical complication in burn patients, marked by rapid and widespread inflammation of the lungs, leading to hypoxemia. ARDS poses significant challenges in clinical management, with mortality rates ranging from 31 to 42.9% in burn patients1,2. The complexity of ARDS in burn patients is further exacerbated by the systemic inflammatory response triggered by extensive tissue damage, which can severely impact respiratory function and complicate treatment approaches3. Despite extensive research on ARDS, the heterogeneity of its presentation in burn patients remains underexplored and insufficiently characterized. Several studies have explored ARDS in burn patients, focusing on diagnostic criteria and mortality rates. For instance, Sine et al. compared the Berlin and American-European Consensus definitions of ARDS in burn patients, underscoring diagnostic challenges unique to this population; however, they did not address the issue of patient heterogeneity2. Wang et al. conducted a systematic review and meta-analysis on the incidence and mortality of ARDS in burn patients, emphasizing the critical nature of the condition without delving into subgroup differences1. In the broader context of ARDS, several studies have identified subphenotypes, emphasizing the importance of understanding patient heterogeneity to enhance treatment outcomes. Wilson and Calfee discussed ARDS subphenotypes in general Intensive Care Unit (ICU) populations, suggesting that heterogeneity in ARDS is a significant issue affecting patient management and outcomes4. Famous et al. similarly identified ARDS subphenotypes with varying responses to fluid management strategies, further emphasizing the necessity of tailored therapeutic approaches5. However, these studies primarily focused on non-burn populations, leaving a gap regarding the heterogeneity of ARDS among burn patients. Research employing clustering techniques has provided further insights into the heterogeneity of ARDS. Bos et al. identified distinct biological phenotypes in ARDS patients through cluster analysis, demonstrating that patient responses to treatments can vary significantly based on underlying biological differences6. However, these studies primarily focused on general ARDS populations, leaving a gap in understanding ARDS among burn patients. Burn patients present unique challenges due to factors such as the extent of total body surface area (TBSA) burned, inhalation injuries, and the systemic inflammatory response—each of which can affect the development and progression of ARDS7. Kim et al. recently employed unsupervised clustering to evaluate clinical heterogeneity and predict mortality in severely burned patients, highlighting the complexity of effectively managing these patients7. However, their study focused primarily on mortality rather than ARDS-specific outcomes and did not investigate the heterogeneity of ARDS presentations in burn patients. The existing literature indicates a need for studies that specifically address the heterogeneity of ARDS in burn patients. Understanding this heterogeneity is crucial for developing personalized treatment strategies and improving patient outcomes. Precision medicine focuses on tailoring healthcare decisions and treatments to individual patients by accounting for variability in genes, environment, and lifestyle8. Applying precision medicine principles to ARDS in burn patients requires a detailed exploration of patient subgroups based on ARDS presentations and outcomes. To address this gap, our study utilizes longitudinal k-means clustering to explore the heterogeneity of ARDS among ICU burn patients. By analyzing time-series data of the PaO2/FiO2 (PF) ratio, a critical indicator of respiratory function, we aim to identify distinct patient subgroups with unique clinical trajectories. We also assess the predictive efficacy of various biomarkers in comparison to the PF ratio within these subgroups. Our focus on a large cohort of burn patients admitted to a specialized burn center over a 12-year period allows us to capture a comprehensive range of ARDS presentations and outcomes, providing robust data for analysis. By identifying distinct clusters within the burn patient population based on ARDS presentations and outcomes, as well as determining predictive biomarkers within these clusters, our study addresses a critical gap in the current literature. This approach aligns with the principles of precision medicine and has the potential to guide more personalized treatment strategies in the ICU, ultimately improving patient outcomes. Methods Study design and participants A retrospective cohort study was conducted in the Burn Intensive Care Unit (BICU) at Hangang Sacred Heart Hospital, affiliated with Hallym University Medical Center in Seoul, Korea. The study period, extending from July 2010 to December 2022, was selected to maximize the sample size and thereby enhance statistical power for detecting ARDS heterogeneity and subgroup effects. The extended timeframe also helped mitigate recruitment challenges typical of rare conditions such as ARDS in burn patients and allowed for multiple rounds of data collection, offering a comprehensive view of patient trajectories and treatment outcomes9,10. The initial cohort comprised 2,654 adults. Exclusion criteria were applied to patients younger than 18 years, those without burn injuries, individuals lacking consistent PF ratio measurements over 48 h, and those with missing essential clinical data upon admission. After these exclusions, the final study cohort consisted of 2,318 patients. Data collection and outcome measures Data were extracted from a prospectively maintained Clinical Data Warehouse (CDW), including demographics, diagnoses, and ICU length of stay. Thirteen clinical variables were collected, including White Blood Cell (WBC) count, Red Cell Distribution Width (RDW), platelet count, Blood Urea Nitrogen (BUN), creatinine, albumin, Prothrombin Time (PT), Total Bilirubin (TB), pH, lactate, partial pressures of oxygen (pO2) and carbon dioxide (pCO2), and the PF ratio. For patients who did not develop ARDS during their ICU stay, laboratory results at ICU admission were used as baseline measurements. For patients who developed ARDS, laboratory results from the time of ARDS diagnosis were used. If multiple laboratory tests were performed on the same day, the most abnormal result was selected to capture the patient’s most severe physiological state. Severity scores, including APACHE IV11, Sequential Organ Failure Assessment (SOFA)12, Abbreviated Burn Severity Index (ABSI)13, revised Baux Score (rBaux)14, and the Hangang score15 were also collected. The primary objective of the study was to assess ARDS incidence, focusing on 60-day in-hospital mortality. PF ratio measurements were retrospectively extracted from medical records for the entire ICU stay, from admission to discharge, to facilitate subgroup identification. ARDS was diagnosed according to the Berlin definition16, with onset identified within one week of known trauma or the onset of new or worsening respiratory symptoms. Chest imaging showed bilateral opacities not fully explained by effusions, collapse, or nodules, ensuring that the observed respiratory failure was not solely attributable to heart failure or fluid overload. Severity was assessed based on the PF ratio, using Positive End-Expiratory Pressure (PEEP) or Continuous Positive Airway Pressure (CPAP) to classify patients into mild, moderate, or severe categories. Acute Kidney Injury (AKI) was defined according to the Acute Kidney Injury Network (AKIN) criteria17, and sepsis was diagnosed using the sepsis-3 criteria18. Data management and statistical analysis Data management adhered to good clinical practice (GCP) guidelines and data protection regulations. Patient identifiers were anonymized to ensure confidentiality. Data completeness and consistency were reviewed, and any discrepancies were resolved by cross-referencing original documents. Winsorization was applied to reduce the influence of outliers. Baseline demographic data were reported as mean ± standard deviation for normally distributed variables and median (interquartile range, IQR) for non-normally distributed variables. Depending on the normality of the data, paired t-tests or Wilcoxon signed-rank tests were employed for continuous variables, while chi-square or Fisher’s exact tests were used for categorical variables19. All statistical analyses were performed using R statistical software (version 3.4.2, R Foundation for Statistical Computing, Vienna, Austria). We adopted a longitudinal k-means (KML) clustering approach for the analysis of time-series data, specifically designed to identify distinct patient subgroups based on their longitudinal PF ratio profiles. The PF ratio was chosen as the sole clustering variable due to its critical role in the clinical diagnosis of ARDS and its widespread use as a key indicator of respiratory function. The clustering algorithm utilized the Euclidean distance metric, which is commonly selected for its straightforward geometric interpretation. The KML approach employed an iterative process to refine cluster centroids and ensure convergence to a stable solution. To address missing data, mean imputation was applied, preserving the overall distribution of the dataset. The calculation of cluster centers was adjusted to account for missing values. Data standardization was an integral part of the process, crucial for normalizing the influence of variables with different scales and variances. This analysis was conducted using the ‘lcMethodKML’ function from the ‘latrend’ package (version 1.5.1) in R20, specialized tool for longitudinal data analysis. The optimal number of clusters for ARDS subgroups was determined using the manual elbow method, guided by key statistical metrics. These included the Consistent Akaike Information Criterion (CAIC), Bayesian Information Criterion (BIC), Classification Likelihood Criterion (CLC), and relative entropy. By plotting these metrics for a range of cluster solutions, from 2 to 10, using the plotMetric function from the latrend package, we could visually inspect the point at which increasing the number of clusters ceased to yield significant improvements. This elbow point, discerned manually from the plot, indicated the optimal balance between the model’s complexity and its fit to the data (Figure S1). We implemented a multivariable logistic regression analysis to identify significant clinical variables of ARDS development, factoring in variables such as age, sex, TBSA, ARDS onset, sepsis, continuous renal replacement therapy (CRRT), and ventilator use. This analysis was instrumental in calculating adjusted odds ratios for these variables. For evaluating the diagnostic accuracy of ARDS, we used a multivariable logistic regression model, assessing its performance with several statistical metrics. These included the AUC, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. The AUC was a central measure of the model’s ability to distinguish between outcomes, with higher values reflecting better performance. The net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were employed to assess improvements in risk classification and discrimination introduced by new clinical variables21. The optimal cutoff point for the diagnostic measures was determined using the Youden index, which maximizes the model’s ability to balance true positive and true negative rates. This selection was pivotal in enhancing the clinical applicability and relevance of the findings. Finally, decision curve analysis was performed to assess the clinical utility of our models, particularly emphasizing variables that demonstrated notable improvements in the NRI and IDI evaluations. This analysis was key to balancing the benefits of true positive diagnoses against the potential harm of false positives, thereby underlining the practical value of the predictive models in a clinical setting. The study was approved by the Institutional Review Board (IRB) of Hangang Sacred Heart Hospital (HG 2024-002), Seoul, Korea. Given the retrospective nature of our study, informed consent was waived by the Institutional Review Board (IRB) of Hangang Sacred Heart Hospital in accordance with national regulations. All methods were conducted in accordance with relevant guidelines and regulations, including the Declaration of Helsinki and the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. Results Overall demographics and cluster distribution The initial cohort included 2,654 ICU patients, of which 2,318 were included in the final analysis after excluding cases with incomplete PF ratio data and missing clinical variables. Patients were divided into four clusters: Cluster A (535 patients), Cluster B (629 patients), Cluster C (666 patients), and Cluster D (488 patients) (Fig. 1). The demographic summary revealed a predominance of male patients (79.3%), with a mean age of 52.5 years. The median TBSA was 34.3%, and the mean length of hospital stay was 22 days. Common complications included sepsis (72.9%) and acute kidney injury (AKI) (59.3%), with 55.7% of patients requiring mechanical ventilation. The overall mortality rate was 24.5%, and the mean APACHE IV and SOFA scores were 48.5 and 4.4, respectively (Table 1). Patient selection and clustering process. Table 1Overall demographic characteristics for enrolled patients.GroupCharacteristicN = 2,3181OutcomesMortality567 (24.5%)LOS22.0 ± 24.9 (1.0–338.0)Onset Day6.7 ± 6.3 (1.0–54.0)Patients’ FactorsAge52.5 ± 16.3 (18.0–99.0)SexMale1,839 (79.3%)Female479 (20.7%)TBSA34.3 ± 24.2 (0.0–100.0)Inhalation903 (39.0%)TypeFB1,668 (72.0%)SB216 (9.3%)EB275 (11.9%)ChB35 (1.5%)CoB124 (5.3%)Severity ScoresABSI8.4 ± 2.7 (2.0–18.0)rBaux93.5 ± 29.0 (23.0–201.0)Hangang133.2 ± 20.7 (94.0–214.0)APACHE IV48.5 ± 29.1 (0.0–172.0)SOFA4.4 ± 3.1 (0.0–19.0)ComplicationsSepsis1,689 (72.9%)ARDS1,015 (43.8%)AKI1,374 (59.3%)InterventionsVentilator1,290 (55.7%)CRRT392 (16.9%)1n (%); Mean ± SD (Range) Overall demographic characteristics for enrolled patients. 1n (%); Mean ± SD (Range) Cluster profiles and characteristics The cohort was divided into four distinct clusters based on clinical characteristics (Fig. 2; Table 2). Cluster A had a mean age of 55.8 years, the highest TBSA (44.2%), and a mortality rate of 54.9%. Sepsis was present in 76.9% of patients, and ARDS developed in 66.1%. Cluster B had the youngest mean age (50.7 years) and a TBSA of 31.9%, with a mortality rate of 18.9%. Sepsis and ARDS occurred in 69.4% and 36.7% of patients, respectively. Cluster C had a mean age of 52.9 years, a TBSA of 32.8%, and a mortality rate of 16.0%. Sepsis was observed in 80.0% of patients, and ARDS in 50.4%. Cluster D had a mean age of 50.9 years and the lowest TBSA (28.4%), with a mortality rate of 7.9%. Sepsis occurred in 65.3% of patients, while ARDS was present in 21.7%. Specific PF ratio values at particular time points and the overall patient trajectory lines are presented in Table S1 and Figure S2. Heatmap and boxplots showcasing the distribution of clinical characteristics and PF ratios respectively, across patient clusters. Table 2Overview of demographic characteristics by cluster.GroupCharacteristicA, N = 5671B, N = 6561C, N = 5641D, N = 5311p-value2OutcomesMortality311 (54.9%)124 (18.9%)90 (16.0%)42 (7.9%)< 0.001LOS20.2 ± 23.4 (2.0–184.0)20.5 ± 22.5 (1.0–272.0)26.8 ± 28.8 (2.0–338.0)20.5 ± 24.1 (2.0–233.0)< 0.001Onset Day5.4 ± 5.4 (1.0–33.0)7.7 ± 5.1 (1.0–30.0)6.1 ± 6.5 (1.0–50.0)10.1 ± 8.7 (1.0–54.0)< 0.001Patients’ FactorsAge55.8 ± 15.6 (19.0–95.0)50.7 ± 16.8 (18.0–96.0)52.9 ± 15.7 (20.0–99.0)50.9 ± 16.4 (18.0–94.0)< 0.001Sex0.027Male475 (83.8%)514 (78.4%)437 (77.5%)413 (77.8%)Female92 (16.2%)142 (21.6%)127 (22.5%)118 (22.2%)TBSA44.2 ± 28.9 (0.0–100.0)31.9 ± 21.8 (0.0–100.0)32.8 ± 22.5 (0.0–100.0)28.4 ± 19.9 (0.0–93.0)< 0.001Inhalation294 (51.9%)244 (37.2%)233 (41.3%)132 (24.9%)< 0.001Type0.006FB441 (77.8%)468 (71.3%)406 (72.0%)353 (66.5%)SB42 (7.4%)54 (8.2%)60 (10.6%)60 (11.3%)EB52 (9.2%)90 (13.7%)56 (9.9%)77 (14.5%)ChB11 (1.9%)9 (1.4%)6 (1.1%)9 (1.7%)CoB21 (3.7%)35 (5.3%)36 (6.4%)32 (6.0%)Severity ScoresABSI9.6 ± 3.1 (4.0–17.0)8.0 ± 2.5 (2.0–16.0)8.3 ± 2.5 (4.0–18.0)7.6 ± 2.1 (3.0–16.0)< 0.001rBaux108.8 ± 32.9 (35.0–197.0)89.0 ± 26.8 (23.0–181.0)92.7 ± 26.3 (24.0–201.0)83.5 ± 22.8 (30.0–160.0)< 0.001Hangang145.9 ± 25.2 (94.0–210.0)128.9 ± 17.6 (94.0–201.0)130.5 ± 17.7 (97.0–214.0)127.6 ± 15.4 (94.0–185.0)< 0.001APACHE IV62.3 ± 32.3 (2.0–158.0)43.2 ± 27.2 (0.0–172.0)47.8 ± 26.0 (2.0–166.0)41.4 ± 25.8 (1.0–134.0)< 0.001SOFA6.3 ± 3.5 (0.0–19.0)3.6 ± 2.8 (0.0–19.0)4.5 ± 2.5 (0.0–15.0)3.2 ± 2.6 (0.0–17.0)< 0.001ComplicationsSepsis436 (76.9%)455 (69.4%)451 (80.0%)347 (65.3%)< 0.001ARDS375 (66.1%)241 (36.7%)284 (50.4%)115 (21.7%)< 0.001AKI446 (78.7%)337 (51.4%)346 (61.3%)245 (46.1%)< 0.001InterventionsVentilator440 (77.6%)309 (47.1%)336 (59.6%)205 (38.6%)< 0.001CRRT176 (31.0%)95 (14.5%)75 (13.3%)46 (8.7%)< 0.0011n (%); Mean ± SD (Range)2Pearson’s Chi-squared test; One-way ANOVA Overview of demographic characteristics by cluster. 1n (%); Mean ± SD (Range) 2Pearson’s Chi-squared test; One-way ANOVA Predictive biomarkers for ARDS across clusters Multivariable logistic regression was performed to analyze the odds ratios of various clinical variables for ARDS diagnosis, and diagnostic accuracy was evaluated using the AUC analysis. In Cluster A, significant clinical variables (p < 0.001) included albumin, WBC count, platelet count, PT, PF ratio, lactate, pO2, pCO2, and RDW. In Cluster B, PF ratio, Albumin, WBC, pO2, Platelet, Lactate, PT, BUN, Creatinine, and RDW were significant. In Cluster C, significant variables included Albumin, WBC, Platelet, RDW, PT, Lactate, Creatinine, BUN, and PF ratio. In Cluster D, significant variables included Albumin, Lactate, WBC, PF ratio, RDW, PT, pO2, Platelet, and Creatinine (Fig. 3, Table S2). AUC analysis demonstrated that albumin had significantly higher predictive power than the PF ratio in Clusters A and C. Detailed AUC values for all clinical variables are provided in Table 3. Adjusted odds ratios with 95% confidence intervals for selected predictors across clusters. Table 3Diagnostic performance and predictive validity of biomarkers for AKI Across patient clusters.ClustersPredictorsAUC (95%CI)p-valueCutoffAccuracySensitivitySpecificityPPVNPVCluster AAlbumin0.905 (0.878 ~ 0.932)0.043 *2.10.8290.8110.8650.3000.079WBC0.895 (0.866 ~ 0.924)0.16021.10.8240.8370.7970.2850.110Platelet0.895 (0.866 ~ 0.923)0.098302.00.8170.8080.8330.3100.096PT0.892 (0.862 ~ 0.922)0.34612.30.8470.9010.7470.1980.132Lactate0.884 (0.855 ~ 0.914)0.6234.10.8120.8350.7670.2930.126PF ratio0.878 (0.846 ~ 0.911)Reference168.60.8610.9250.7340.1660.128pO20.872 (0.839 ~ 0.906)0.48163.90.8360.8800.7490.2390.127RDW0.869 (0.836 ~ 0.901)0.36012.60.8270.8720.7400.2530.133pCO20.869 (0.836 ~ 0.901)0.32739.90.8290.8880.7120.2360.142Cluster BPF ratio0.960 (0.946 ~ 0.973)Reference259.70.8870.9540.8480.0300.215Albumin0.944 (0.926 ~ 0.961)0.0522.70.8770.9040.8610.0610.208WBC0.940 (0.923 ~ 0.958)0.023 *21.10.8710.8290.8950.1000.178pO20.936 (0.917 ~ 0.954)< 0.001 **94.10.8670.8670.8670.0820.208PT0.929 (0.909 ~ 0.949)< 0.001 **12.30.8580.8910.8390.0690.242Lactate0.923 (0.903 ~ 0.944)< 0.001 **4.10.8260.9400.7600.0440.306Platelet0.922 (0.901 ~ 0.944)< 0.001 **257.00.8480.8830.8270.0760.251Creatinine0.910 (0.887 ~ 0.934)< 0.001 **0.60.8370.8710.8180.0840.264BUN0.910 (0.887 ~ 0.933)< 0.001 **11.30.8360.8380.8350.1020.253RDW0.908 (0.886 ~ 0.931)< 0.001 **13.00.8280.8870.7930.0760.285pCO20.906 (0.882 ~ 0.929)< 0.001 **52.30.8320.8590.8160.0910.269Cluster CAlbumin0.915 (0.891 ~ 0.938)< 0.001 **3.00.8510.8730.8290.1340.162Platelet0.893 (0.867 ~ 0.920)0.059134.00.8170.8450.7890.1660.197WBC0.893 (0.867 ~ 0.919)0.06615.20.8190.8030.8360.1930.168PT0.890 (0.863 ~ 0.918)0.20912.80.8260.8320.8210.1700.178RDW0.889 (0.863 ~ 0.916)0.14914.70.8190.8490.7890.1630.197Lactate0.887 (0.860 ~ 0.914)0.2571.70.8120.7560.8700.2240.142Creatinine0.885 (0.858 ~ 0.912)0.4040.80.8170.7780.8570.2080.153BUN0.881 (0.853 ~ 0.908)0.8157.10.8100.9150.7040.1090.242pO20.879 (0.851 ~ 0.907)0.94295.30.8170.8170.8180.1850.180PF ratio0.879 (0.851 ~ 0.907)Reference117.30.8140.8200.8070.1840.188Cluster DAlbumin0.938 (0.916 ~ 0.959)0.5513.20.8700.8850.8650.0350.359Lactate0.933 (0.906 ~ 0.961)0.7271.80.8490.9020.8350.0320.399PF ratio0.929 (0.902 ~ 0.955)Reference117.30.8850.8090.9060.0550.295PT0.927 (0.899 ~ 0.955)0.23912.50.9030.8300.9220.0460.267WBC0.926 (0.898 ~ 0.954)0.57114.30.8640.8670.8630.0400.368pO20.919 (0.891 ~ 0.947)0.08872.20.8660.8350.8750.0500.351RDW0.917 (0.886 ~ 0.948)0.10012.70.8680.8320.8770.0490.352Creatinine0.912 (0.883 ~ 0.941)0.008 *0.50.8660.8230.8770.0520.354Platelet0.910 (0.878 ~ 0.942)0.038 *179.00.8900.7790.9210.0610.273BUN0.904 (0.873 ~ 0.935)< 0.001 **13.80.8390.8580.8340.0440.416pH0.904 (0.872 ~ 0.936)< 0.001 **7.430.8430.8230.8490.0540.404pCO20.900 (0.868 ~ 0.933)< 0.001 **46.40.8680.8260.8800.0520.345 Diagnostic performance and predictive validity of biomarkers for AKI Across patient clusters. Reclassification analysis for ARDS prediction across clusters Reclassification analysis showed that albumin in Cluster A improved prediction over the PF ratio, with a NRI of 0.421 and an IDI of 0.033. In Cluster C, Albumin, WBC, Platelet, PT, and RDW were significant, with Albumin showing an NRI of 0.845 and an IDI of 0.096, and Lactate showing an NRI of 0.329. No significant improvements in prediction were found in Cluster B over the PF ratio. In Cluster D, Lactate and WBC had NRIs of 0.293 and 0.219, respectively (Figs. 4 and 5, Tables S3 and S4). Net Reclassification Improvement (NRI) values with 95% confidence intervals. Integrated Discrimination Improvement (IDI) statistics across clusters. Biomarker effectiveness in ARDS prediction: decision curve analysis Decision curve analysis showed that albumin in Cluster A provided higher net benefits at thresholds above 70%. In Cluster B, the PF ratio remained the most predictive marker. In Cluster C, although albumin and other biomarkers showed potential usefulness, they did not significantly outperform the PF ratio. In Cluster D, lactate and WBC demonstrated better performance at higher thresholds (Fig. 6, Table S5). Decision curves highlight the clinical benefits of different predictive models based on patient clusters. Discussions Our study identified four distinct clusters of ICU burn patients using longitudinal PF ratio trajectories, highlighting significant heterogeneity in ARDS progression within this population. This heterogeneity reflects the complex interplay of factors such as burn severity, inhalation injury, and complications like sepsis and AKI, which are known to influence ARDS development in burn patients. Cluster A patients had the highest incidence of ARDS (66.1%) and mortality rate (54.9%), along with the largest average TBSA burned (44.2%) and the highest prevalence of inhalation injury (51.9%). These findings are consistent with previous studies indicating that larger burn size and inhalation injury are significant risk factors for ARDS in burn patients1,2. The severe physiological stress and systemic inflammatory response in this group likely contribute to worse outcomes. In contrast, Cluster D had the lowest incidence of ARDS (21.7%) and mortality rate (7.9%), indicating a less severe clinical trajectory. The variability among clusters underscores the importance of recognizing different patient subgroups within the burn population, as their risk profiles and clinical courses differ significantly. Albumin emerged as a significant predictor of ARDS in Clusters A and C, with AUC values of 0.905 and 0.915, respectively. Hypoalbuminemia is common in burn patients due to increased vascular permeability and protein loss from burn wounds, which reduces oncotic pressure and contributes to pulmonary edema and ARDS3,22. The strong association between low Albumin levels and ARDS in these clusters suggests that Albumin could serve as a valuable biomarker for early identification of patients at higher risk for ARDS. This finding aligns with recent studies that have highlighted the prognostic value of Albumin levels in critically ill patients, including those with ARDS3,23. However, the inconsistent predictive value of Albumin across all clusters indicates that its utility may be limited to specific patient subgroups. In Clusters B and D, biomarkers did not significantly enhance predictive power over the PF ratio, which remained a robust predictor of ARDS. This suggests that while biomarkers like Albumin can be helpful in certain subgroups, traditional clinical metrics such as the PF ratio are still essential for ARDS prediction in burn patients. Combining biomarkers with clinical assessments may improve early detection and enable personalized treatment strategies, aligning with the principles of precision medicine8,24. Our findings expand on previous research that emphasized the heterogeneity of ARDS in burn patients and the importance of identifying subphenotypes for targeted interventions5,6. By incorporating biomarker profiles and longitudinal data analysis, we provide a deeper understanding of ARDS subtypes beyond traditional classifications. A strength of our study is the use of longitudinal k-means clustering, which captures the dynamic progression of ARDS and allows for the identification of distinct patient subgroups based on PF ratio trajectories. This approach has been effective in other studies for identifying biological phenotypes in ARDS patients6,25. By analyzing temporal trends, we can better understand disease progression and tailor interventions accordingly. However, several limitations should be acknowledged. The retrospective design and single-center setting may limit the generalizability of our findings. Additionally, variability in the timing of biomarker measurements could have impacted predictive accuracy, despite efforts to standardize data collection points. Moreover, the underlying mechanisms linking hypoalbuminemia to ARDS development in burn patients were not explored in this study, warranting further investigation. Recognizing the heterogeneity of ARDS among burn patients is crucial for optimizing management strategies. Integrating biomarkers like Albumin with traditional clinical assessments could enhance early identification of high-risk patients, particularly in more severe cases represented by Cluster A. This could facilitate timely interventions, such as aggressive monitoring, fluid management, and nutritional support, potentially improving outcomes. Our study underscores the importance of a multifaceted approach in ARDS management among burn patients, combining clinical metrics and biomarkers to guide personalized treatment strategies. This approach could improve risk stratification and therapeutic decisions, ultimately enhancing patient care. Further research should focus on validating the predictive value of Albumin and other biomarkers in larger, multicenter cohorts of burn patients. Developing comprehensive predictive models that incorporate biomarkers, genetic, and clinical variables may enhance precision medicine approaches in ARDS care25. Additionally, investigating how burn injuries modulate biomarker levels could reveal new therapeutic targets and inform earlier intervention strategies. Conclusions Our study highlights significant heterogeneity in ARDS progression among ICU burn patients, identifying distinct clusters with varying risks and outcomes. Albumin was a strong predictor of ARDS in specific patient subgroups, particularly those with more severe conditions. However, its predictive value was not consistent across all clusters, indicating that reliance on a single biomarker is insufficient for accurate ARDS prediction in burn patients. By combining biomarkers like Albumin with traditional clinical metrics such as the PF ratio, we may improve early detection of ARDS and tailor personalized treatment strategies. Recognizing the variability among patient subgroups is essential for optimizing ARDS management in burn patients. Further prospective, multicenter studies are required to validate these findings and develop comprehensive predictive models to improve patient outcomes. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Abbreviations Acute respiratory distress syndrome Intensive Care Unit Total body surface area PaO2/FiO2 White blood cell Red cell distribution width Blood urea nitrogen Prothrombin time Total bilirubin Acute kidney injury acute physiology and chronic health evaluation IV Sequential organ failure assessment Abbreviated burn severity index Revised Baux score Continuous renal replacement therapy Positive end-expiratory pressure Continuous positive airway pressure Acute kidney injury network Area under the curve Net reclassification improvement Integrated discrimination improvement Author contributions JY curated data, led the investigation, and managed project administration. DK conceptualized the research, developed software, supervised the project, and wrote the original draft of the manuscript. YSC curated data, validated the findings, and created visual representations. JH contributed to the investigation, developed the methodology, supervised the project, and created visualizations. DY conducted formal analysis, developed software, and contributed to reviewing and editing the manuscript. All authors reviewed and approved the final manuscript. Funding No funding was received for this study. Data availability The datasets generated and/or analyzed during this study are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Ethics approval and consent to participate The study was approved by the Institutional Review Board (IRB) of Hangang Sacred Heart Hospital (HG 2024-002), Seoul, Korea. Given the retrospective nature of our study, informed consent was waived by the Institutional Review Board (IRB) of Hangang Sacred Heart Hospital in accordance with national regulations. All methods were performed in accordance with the relevant guidelines and regulations, including the Declaration of Helsinki and the STROBE (Strengthening the Reporting of Observational Studies in Epidemiology) guidelines. Consent for publication Not applicable. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The datasets generated and/or analyzed during this study are available from the corresponding author on reasonable request.