Modelagem por equações estruturais aplicada à predição de sepse em dados clínicos multicêntricos

Abstract

Sepsis is a complex clinical syndrome associated with high global mortality, whose early diagnosis remains a major challenge, especially in resource-limited settings. Computational models based on machine learning have shown good predictive performance but lack transparency and explanatory capacity. In this context, Structural Equation Modeling (SEM) emerges as a methodological alternative for representing latent clinical constructs and testing the causal relationships underlying sepsis risk. The present study aimed to construct and evaluate a theoretical explanatory model using SEM, based on data from the PhysioNet/Computing in Cardiology Challenge 2019. Initially, data preparation and integration steps were carried out, including format conversion, concatenation, selection of clinical variables, and statistical standardization. Subsequently, latent constructs representing physiopathological systems were specified, and the model was estimated using the semopy package (Python). This approach enabled the calculation of direct and indirect effects, the assessment of global fit quality using the Comparative Fit Index (CFI), Tucker-Lewis Index (TLI), and Root Mean Square Error of Approximation (RMSEA), Goodness of Fit Index (GFI), Adjusted Goodness of Fit Index (AGFI), and the generation of graphical representations of clinical trajectories. The General Condition construct had the greatest impact (standardized coefficient = 0.123; p < 0.001), explaining approximately 30 % of the model’s variance. Next, Respiratory Risk showed a positive contribution of moderate magnitude (standardized coefficient = 0.041; p < 0.001), accounting for about 9 % of the relative influence. In contrast, the Cardiovascular Risk construct showed a negative association of low magnitude (standardized coefficient = - 0.045; p < 0.001), corresponding to approximately 5 % of relative contribution to the outcome. SEM indicated adequate global model fit, with CFI = 0.988, GFI = 0.988, NFI = 0.988, AGFI = 0.942, TLI = 0.942, and a low approximation error (RMSEA = 0.025), confirming the statistical robustness of the estimated relationships. It is concluded that SEM applies to the clinical context of sepsis, contributing not only to prediction but also to the causal explanation of physiopathological mechanisms, thus representing a promising tool for integrating statistical precision and clinical relevance in decision-support systems.