Prognostic effects of glycaemic variability on diastolic heart failure and type 2 diabetes mellitus: insights and 1-year mortality machine learning prediction model

Abstract

Abstract Background Diastolic heart failure (DHF) and type 2 diabetes mellitus (T2DM) often coexist, causing increased mortality rates. Glycaemic variability (GV) exacerbates cardiovascular complications, but its impact on outcomes in patients with DHF and T2DM remains unclear. This study examined the relationships between GV with mortality outcomes, and developed a machine learning (ML) model for long-term mortality in these patients. Methods Patients with DHF and T2DM were included from the Medical Information Mart for Intensive Care IV, with admissions (2008–2019) as primary analysis cohort and admissions (2020–2022) as external validation cohort. Multivariate Cox proportional hazards models and restricted cubic spline analyses were used to evaluate the associations of GV with 90-day, 1-year, and 3-year all-cause mortality. The primary analysis cohort was split into training and internal validation cohorts, then developing ML models for predicting 1-year all-cause mortality in training cohort, which were validated using the internal and external validation cohorts. Results 2,128 patients with DHF and T2DM were included in primary analysis cohort (meidian age 71.0years [IQR: 62.0–79.0]; 46.9% male), 498 patients with DHF and T2DM were included in the external validation cohort (meidian age 75.0years [IQR: 67.0–81.0]; 54.0% male). Multivariate Cox proportional hazards models showed that high GV tertiles were associated with higher risk of 90-day (T2: HR 1.45, 95%CI 1.09–1.93; T3: HR 1.96, 95%CI 1.48–2.60), 1-year (T2: HR 1.25, 95%CI 1.02–1.53; T3: HR 1.54, 95%CI 1.26–1.89), and 3-year (T2: HR 1.31, 95%CI: 1.10–1.56; T3: HR 1.48, 95%CI 1.23–1.77) all-cause mortality, compared with lowest GV tertile. Chronic kidney disease, creatinine, potassium, haemoglobin, and white blood cell were identified as mediators of GV and 1-year all-cause mortality. Additionally, GV and other clinical features were pre-selected to construct ML models. The random forest model performed best, with AUC (0.770) and G-mean (0.591) in internal validation, with AUC (0.753) and G-mean (0.599) in external validation. Conclusion GV was determined as an independent risk factor for short-term and long-term all-cause mortality in patients with DHF and T2DM, with a potential intervention threshold around 25.0%. The ML model incorporating GV demonstrated strong predictive performance for 1-year all-cause mortality, highlighting its importance in early risk stratification management of these patients.

Keywords

• Glycaemic variability
• Heart failure with preserved ejection fraction
• Type 2 diabetes mellitus
• Mortality
• Machine learning