Dynamic HGI trajectories and their impact on survival in patients with sepsis: a machine learning prognostic model

Previous studies have indicated a correlation between the glycosylated hemoglobin index (HGI) and the prognosis of patients with sepsis. However, the impact of its dynamic fluctuations on patient outcomes remains insufficiently explored. This study seeks to investigate the relationship between HGI trajectory changes over time and the prognosis of patients with sepsis, and developing an optimal predictive model for 28-day and 90-day mortality risk using machine learning techniques.

Data from the MIMIC-IV 3.0 database were employed to construct a linear regression model utilizing glycosylated hemoglobin (HbA1c) and fasting blood glucose (FBG) measured at six distinct time points to calculate HGI. Trajectory analysis revealed three distinct HGI change patterns: rising (RS), stable (ST), and descending (DS). Based on these groupings, Kaplan-Meier survival curves, Cox regression models, and mediation analysis were applied. Lasso regression was utilized for feature selection, and four machine learning models-Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), Logistic Regression, and Random Forest (RF)-were developed and evaluated through ROC curves, Decision Curve Analysis (DCA), and calibration curves. SHapley Additive exPlanations (SHAP) values were incorporated for interpretability, and nomograms for 28-day and 90-day mortality were generated.

A total of 2,616 patients were included in the analysis, with 407 patients (15.56%) dying within 28 days. multivariable Cox regression analysis, using the ST group as a reference, revealed that patients in the RS group had a significantly higher risk of death at both 28 days and 90 days, whereas those in the DS group demonstrated a markedly reduced risk. These findings were consistent across all models. Among the models evaluated, the Logistic Regression model exhibited the highest predictive accuracy, with AUC values for the 28-day mortality prediction in the training and validation sets of 0.743 and 0.732, respectively, and for 90-day mortality of 0.748 and 0.735. The key predictive factors for 28-day mortality included albumin, HGI, renal injury, and the APSIII score, while for 90-day mortality, renal injury, APSIII score, HGI, and albumin were the most influential variables. To enhance the clinical applicability of these models, nomograms for 28-day and 90-day mortality were constructed, achieving AUCs of 0.715 and 0.747, respectively. Calibration curves demonstrated strong concordance between predicted probabilities and observed outcomes.

Dynamic alterations in HGI during follow-up were found to be significantly associated with the risk of mortality in patients with sepsis at both 28 days and 90 days. This study presents a machine learning-based model for predicting the mortality risk of patients with sepsis, with potential clinical utility in early risk stratification and improving patient prognosis.