Parenclitic network mapping predicts survival in critically ill patients with sepsis

Abstract

Sepsis is a complex disease involving multiple organ systems. A network physiology approach to sepsis may reveal collective system behaviors and intrinsic organ interactions. However, mapping functional connectivity for individual patients has been challenging due to the lack of analytical methods for evaluating physiological networks using routine clinical and laboratory data. This study explored the use of parenclitic network mapping to assess organ connectivity and predict sepsis outcomes based on routine laboratory data. Data from 162 sepsis patients meeting Sepsis-3 criteria were retrospectively analyzed from the MIMIC-III database. Fifteen physiological variables representing organ systems were used to construct organ network connectivity through correlation analysis. Correlation analysis identified 7 interactions linked to 30-day survival. Parenclitic network analysis was used to measure deviations in individual patients' correlations between organ systems from the reference physiological interactions observed in survivors. Parenclitic deviations in the pH-bicarbonate axis (hazard ratio = 2.081, p < 0.001) and pH-lactate axis (hazard ratio = 2.773, p = 0.024) significantly predicted 30-day mortality, independent of the Sequential Organ Failure Assessment (SOFA) score and ventilation status. This study highlights the potential of parenclitic network mapping to provide insights into sepsis pathophysiology and differences in organ system connectivity between survivors and non-survivors independent of sepsis severity and mechanical ventilation status.

Keywords: intensive care; network; network physiology; parenclitic; sepsis; survival.

FIGURE 1

(a) Diagram illustrating the differences between linear and non‐linear models of sepsis pathophysiology. (b) Diagram illustrating different network physiological techniques used for population‐level (A) and individual‐level (B) analysis of organ network connectivity.

FIGURE 2

(a) An overview of the research design. (b) The flowchart illustrates the complete data extraction process from the MIMIC‐III database. The purple boxes show the patient data included in the present work, investigating the application of static network mapping (parenclitic analysis) in predicting sepsis outcomes. The yellow boxes highlight the study population included in the earlier work by Morandotti et al. (2025), which explored the use of dynamic network mapping (transfer entropy) in prognosticating sepsis outcomes. (c) The figure illustrates 15 physiological variables representing different organ systems. Each variable is categorized according to the systems that it represents. Variables belonging to more than one category may reflect the function of multiple physiological systems. (d) Process behind parenclitic network mapping. (1) Population correlation map generation. (2) Computation of individual parenclitic network for significant physiological correlations identified from population correlation maps. “r” represents the correlation coefficient for the regression line of significantly correlated physiological variables (i.e., A and B) within the reference population (i.e., survivors). “” represents the parenclitic deviation (PD). “∆” represents the difference between the reference population and individual patient data. All PD computations were conducted using in‐house code developed in MATLAB.

FIGURE 3

(a) Population Correlation Network Maps for 30‐day survival (A) and 48‐h deterioration (B). Nodes represent physiological variables. Edges show significant correlations (p‐value ≤Bonferroni‐adjusted p‐value) between the two nodes. Edge labels display the values of Pearson's correlation coefficient. (b) Kaplan–Meier Graphs illustrating 30‐day survival of patients classified into predicted survivor and non‐survivor groups based on (A) PD (pH – HCO₃ ⁻), (B) PD (lactate – pH), and (C) SOFA score cut‐offs.