Rapid Sepsis Detection Using Routine Laboratory Data: A Logistic Regression Approach

Abstract

Background
Sepsis is a life-threatening condition associated with high morbidity and mortality, in which early diagnosis is crucial for improving clinical outcomes [1] Routine haematological parameters generated by automated analyzers have emerged as promising biomarkers for sepsis detection due to their rapid availability and low cost [2].

Aims
The principal aim was to investigate the predictive value of routine hematochemical parameters for sepsis diagnosis and to construct a validated multivariable logistic regression model. Particular attention was given to variable selection, model validation, and cut-off optimization through the weighted Youden index to maximize sensitivity while maintaining high specificity.

Methods
A total of 597 subjects admitted to the Emergency Department were included and the sample was randomly divided into training (n = 427; 71.5%) and validation datasets (n = 170; 28.5%). Dataset homogeneity was assessed through independent samples t-tests and chi-square tests with effect size estimation (Cohen’s d and Cramer’s V). Multicollinearity was evaluated using variance inflation factors (VIF > 10 threshold) and univariate logistic regression analyses were initially performed to identify candidate predictors associated with sepsis, with variable preselection based on statistical significance and sensitivity > 0.70. A forward stepwise multivariable logistic regression approach was subsequently applied (entry p = 0.05; removal p = 0.10). Model fit was evaluated using Nagelkerke’s R² and Akaike Information Criterion (AIC). Discriminative ability was assessed using ROC curves, area under the curve (AUC), sensitivity, specificity, and weighted Youden index (J_w). Since minimizing false negatives represented the main clinical priority, multiple probability thresholds (0.25–0.75) were tested to optimize classification performance.

Results
Most hematochemical parameters differed significantly between sepsis and non-sepsis groups. Nine variables demonstrated strong discriminatory performance in univariate analyses (sensitivity > 0.70). The final multivariable model included ten predictors and demonstrated excellent goodness-of-fit (Nagelkerke’s R² = 0.947; AIC = 84.689). At the conventional threshold (0.50), sensitivity and specificity reached 97.12% and 97.26% in the training dataset, and 92.96% and 96.67% in validation. Threshold optimization identified 0.31 as the optimal cut-off (J_w = 0.9796), increasing sensitivity to 99.04% while maintaining specificity at 96.35%, substantially reducing false negatives.

Conclusions
The proposed statistical framework enabled the development of a highly discriminative and robust predictive model for sepsis detection. The prioritization of sensitivity-driven threshold optimization improved clinical applicability by minimizing missed diagnoses while preserving high specificity.