AdaBoost ML model predicts MACE in dialysis patients with 80.9% AUC using NT-proBNP, eGFR, GLS, and age
Background
Patients undergoing dialysis face a significantly elevated risk of major adverse cardiovascular events (MACE), including myocardial infarction, heart failure, and cardiovascular death. Traditional risk assessment tools often fall short in accurately identifying these high-risk individuals, leading to suboptimal preventative strategies. Developing robust predictive models is crucial for early intervention and improving prognosis in this vulnerable population. This study addresses this gap by leveraging machine learning to integrate diverse clinical and echocardiographic parameters for more precise risk stratification.
Study Design
This retrospective cohort study included 203 dialysis patients (median age 45.0 years, 64.0% male) with an average follow-up of 18 months. Participants were split into training and test sets at a 7:3 ratio. LASSO regression was employed to select key characteristic variables from general information, laboratory tests, and echocardiographic parameters, including global longitudinal strain (GLS). Eight distinct machine learning models were then constructed, and SHAP analysis was used to evaluate the importance of each feature in the best-performing model.
Results
The incidence of MACE (myocardial infarction, unstable angina, heart failure, cardiovascular death) in the dialysis cohort was 38.92%. Among the eight machine learning models developed, AdaBoost demonstrated superior predictive performance. In the test set, AdaBoost achieved an AUC of 0.809 (95% CI: 0.706-0.912), with an accuracy of 0.750, sensitivity of 0.90, and specificity of 0.675. The SHAP analysis identified four primary features as most important for predicting MACE:
N-terminal pro-brain natriuretic peptide (NT-proBNP) level (mean absolute
SHAP0.199), estimated glomerular filtration rate (eGFR) (0.176), global longitudinal strain (GLS) (0.096), and age (0.091). These findings highlight the critical role of cardiac biomarkers, renal function, and cardiac mechanics in MACE prediction.
Key Findings
- MACE incidence in dialysis patients was 38.92% over an 18-month follow-up.
- AdaBoost ML model achieved a test set AUC of 0.809 for MACE prediction.
- AdaBoost model showed 75.0% accuracy, 90% sensitivity, and 67.5% specificity in the test set.
- Elevated NT-proBNP was the strongest predictor of MACE (mean absolute
SHAP0.199). - Reduced eGFR, impaired GLS, and advanced age were also significant MACE predictors.
Why It Matters
This study provides a robust machine learning model that could significantly improve risk stratification for dialysis patients prone to MACE. By integrating readily available clinical and echocardiographic data, the model offers a data-driven approach to identify high-risk individuals more accurately than traditional methods. This enhanced predictive capability could enable clinicians to implement earlier, more targeted interventions, such as intensified monitoring, personalized medication adjustments, or lifestyle modifications, potentially reducing the incidence of severe cardiovascular events. The identification of NT-proBNP, eGFR, GLS, and age as key predictors also reinforces their importance in clinical assessment and could guide future research into specific therapeutic targets.
machine-learning
cardiovascular-events
dialysis
mace
risk-prediction
nt-probnp