Predicting 28-day all-cause mortality in patients admitted to intensive care units with pre-existing chronic heart failure using the stress hyperglycemia ratio: a machine learning-driven retrospective cohort analysis

Abstract

Chronic heart failure (CHF) poses a significant threat to human health. The stress hyperglycemia ratio (SHR) is a novel metric for accurately assessing stress hyperglycemia, which has been correlated with adverse outcomes in various major diseases. However, it remains unclear whether SHR is associated with 28-day mortality in patients with pre-existing CHF who were admitted to intensive care units (ICUs). This study retrospectively recruited patients who were admitted to ICUs with both acute critical illness and pre-existing CHF from the Medical Information Mart for Intensive Care (MIMIC) database. Characteristics were compared between the survival and non-survival groups. The relationship between SHR and 28-day all-cause mortality was analyzed using restricted cubic splines, receiver operating characteristic (ROC) curves, Kaplan-Meier survival analysis, and Cox proportional hazards regression analysis. The importance of the potential risk factors was assessed using the Boruta algorithm. Prediction models were constructed using machine learning algorithms. A total of 913 patients were enrolled. The risk of 28-day mortality increased with higher SHR levels (P < 0.001). SHR was independently associated with 28-day all-cause mortality, with an unadjusted hazard ratio (HR) of 1.45 (P < 0.001) and an adjusted HR of 1.43 (P < 0.001). Subgroup analysis found that none of the potential risk factors, such as demographics, comorbidities, and drugs, affected the relationship (P for interaction > 0.05). The area under the ROC (AUC) curve for SHR was larger than those for admission blood glucose and HbA1c; the cut-off for SHR was 0.57. Patients with SHR higher than the cut-off had a significantly lower 28-day survival probability (P < 0.001). SHR was identified as one of the key factors for 28-day mortality by the Boruta algorithm. The predictive performance was verified through four machine learning algorithms, with the neural network algorithm being the best (AUC 0.801). For patients with both acute critical illness and pre-existing CHF, SHR was an independent predictor of 28-day all-cause mortality. Its prognostic performance surpasses those of HbA1c and blood glucose, and prognostic models based on SHR provide clinicians with an effective tool to make therapeutic decisions.

Keywords: Chronic heart failure; Machine learning; Mortality; Risk factor; Stress hyperglycemia ratio.

Fig. 1

Inclusion flow in this study. CHF Chronic heart failure, HbA1C Glycosylated hemoglobin

Fig. 2

The association between SHR and 28-day mortality by restricted cubic spline method A The unadjusted evaluation; B The adjusted evaluation. CI Confidence interval, SHR Stress hyperglycemia ratio

Fig. 3

COX proportional hazards regression and the subgroup analysis. OASIS Oxford acute illness severity score, HR Hazard ratio

Fig. 4

Prediction performance for 28-day all-cause mortality by ROC curves. SHR Stress hyperglycemia ratio, ABG Admission blood glucose, HbA1C Glycosylated hemoglobin

Fig. 5

Kaplan–Meier survival curve for mortality according to SHR cut-off SHR: stress hyperglycemia ratio

Fig. 6

Importance of potential risk factors of 28-day mortality ranked by Boruta algorithm. The horizontal axis is the name of each variable, and the vertical axis is the Z value of each variable. The box plot shows the Z value of each variable during model calculation. The green boxes represent important variables, the red boxes represent unimportant variables, and the yellow boxes represent potentially important variables. ABG Admission blood glucose, HbA1c Glycosylated hemoglobin, OASIS Oxford acute illness severity score, MAP Mean arterial pressure, SpO₂ Pulse oxygen saturation, SHR Stress hyperglycemia ratio, MAP Mean arterial pressure, PTT Partial thromboplastin time, INR International normalized ratio, PT Prothrombin time, BUN Blood urea nitrogen, HbA1C Glycosylated hemoglobin, AST Aspartate aminotransferase, ALT Alanine aminotransferase, COPD Chronic obstructive pulmonary disease, WBC White blood cell count, SAP Stable angina pectoris, UAP Unstable angina pectoris, AMI: acute myocardial infarction, ARDS Acute respiratory distress disease

Fig. 7

Establishment and validation of the machine learning prediction model. A ROC curve of the machine learning model. B DCA of the machine learning model. C Calibration curve of the KNN algorithm model. KNN K-Nearest Neighbors, Coxph Cox proportional hazards survival