An XGBoost machine learning model predicted 12-month heart failure readmission with AUC 0.876, outperforming logistic regression (AUC 0.859, p=0.016) in 14,131 ED patients.
Does an XGBoost machine learning model improve the prediction of heart failure readmission within 12 months compared to standard logistic regression in undifferentiated ED patients with suspected cardiac conditions?
An XGBoost machine learning model significantly outperforms standard logistic regression in predicting 12-month heart failure readmission among emergency department patients with suspected cardiac conditions.
Absolute Event Rate: 0% vs 0%
Abstract Background Heart failure (HF) remains a major global public health challenge contributing to substantial morbidity and mortality, as well as health system resource utilization. In Australia, HF accounts for approximately 1.5% of all hospital admissions per year and is a leading cause for hospital readmission within 90 days of discharge. Conventional models for predicting HF readmissions are often limited by clinical applicability among patient populations with well recognized risk factors. Purpose This study aimed to apply a machine learning (ML) approach for the prediction of HF readmission within 12 months for undifferentiated patients presenting to emergency departments (EDs) across a broad range of suspected cardiac conditions. Methods This was a sub-analysis of a cluster randomized clinical trial (n=14,131) across 12 EDs in South Australia from April to December 2023 which enrolled patients with symptoms warranting high-sensitivity cardiac troponin T (hs-cTnT) testing and were then followed for at least 12 months for heart failure. The dataset was randomly split into ML model training (70%) and test (30%) sets. Four ML algorithms were applied and compared to standard logistic regression. Features with more than 25% missing data were excluded, resulting in 38 features being included for analysis. Model performance was evaluated by the area under the receiver operating characteristic curve (AUC). For the highest performing ML model, feature importance was evaluated using the SHapley Additive exPlanations (SHAP) method. Results Among 14,131 prospectively identified ED patients (49% female), 704 patients (4.98%) were readmitted with HF. Median time to HF admission was 69.5 days IQR: 25 to 149. In the test set, an Extreme Gradient Boosting (XGBoost) model (AUC 0.876; CI 0.861 - 0.890) outperformed standard logistical regression (AUC 0.859; CI 0.842 - 0.875, p=0.016). For the remaining ML models, AUCs were 0.857, 0.869 and 0.857 for random forest, LASSO and neural networks, respectively. The XGBoost brier sore was 0.041, demonstrating good model calibration. Within the XGBoost model, SHAP values indicated that higher troponin concentrations, older age, pre-existing heart failure, known arrhythmias and a higher Charlson comorbidity index were the highest-ranked predictors of HF readmission. Conclusion A comprehensive XGBoost-based ML model demonstrated high discrimination for the prediction of HF readmission, outperforming conventional logistic regression. This model, if implemented in clinical practice has the potential to identify patients at high-risk of HF and guide early intervention strategies.Machine Learning model AUC curves Shapley Values Feature Importance
Goel et al. (Sat,) reported a other. An XGBoost machine learning model predicted 12-month heart failure readmission with AUC 0.876, outperforming logistic regression (AUC 0.859, p=0.016) in 14,131 ED patients.