Evaluating clinical heterogeneity and predicting mortality in severely burned patients through unsupervised clustering and latent class analysis

Abstract

Burn injuries often result in a high level of clinical heterogeneity and poor prognosis in patients with severe burns. Clustering algorithms, which are unsupervised methods that can identify groups with similar trajectories in patients with heterogeneous diseases, can provide insights into the mechanisms of the disease pathogenesis. This study aimed to analyze routinely collected biomarkers to understand their mortality prediction power, identify the clinical meanings or subtypes, and inform treatment decisions to improve the outcomes of patients with burns. This retrospective cohort study included patients aged ≥ 18 years who were admitted between January 2010 and December 2021. The patients were divided into four subgroups based on the time period of their admission: week 1, 2, 3, and 4. The study revealed that 22 biomarkers were evaluated, and the red blood cell distribution width, bicarbonate level, pH, platelets, and lymphocytes were significantly associated with the mortality risk. Latent class analysis further demonstrated that the pH, platelets, lymphocytes, lactate, and albumin demonstrated the lowest levels in the cluster with the highest risk of mortality, with the lowest levels of pH and lactate being particularly noteworthy in week 1 of the study. During the week 2, the pH and lymphocyte levels were demonstrated to be significant predictors of the mortality risk, whereas the lymphocyte and platelet counts were meaningful predictors in week 3. During week 4, pH, platelet count, and albumin level were important predictors of mortality risk. Analysis of routinely collected biomarkers using clustering algorithms and latent class analysis can provide valuable insights into the heterogeneity of burn injuries and improve the ability to predict disease progression and mortality. Our findings suggest that lactate levels are a better indicator of cellular hypoxia in the early stages of burn shock, whereas platelet and lymphocyte levels are more indicative of infections such as sepsis. Albumin levels are considered a better indicator of reduced nutritional loss with decrease in unhealed burn wounds; however, the pH levels reflect the overall condition of the patient throughout the study period. These findings can be used to inform treatment decisions and improve the outcomes of burn patients.