Development and external validation of a prediction model for prolonged intensive care unit stay in heart failure patients.

AIMS: Prolonged intensive care unit (ICU) stays in heart failure patients are associated with poor prognosis and result in high medical expenses. To develop and validate a predictive model for prolonged ICU stays in heart failure patients.

METHODS AND

RESULTS

A retrospective cohort study was conducted involving ICU patients with a primary diagnosis of heart failure. The model development cohort comprised 5744 ICU heart failure patients from Beth Israel Deaconess Medical Center, while the external validation cohort included 4056 ICU heart failure patients from 208 hospitals across the USA. The primary outcome was prolonged ICU stay in heart failure patients. Feature selection was performed using Least Absolute Shrinkage and Selection Operator, univariate, and multivariate logistic regression. Nine machine learning algorithms were applied to develop the prediction models. Based on discrimination, calibration, and clinical utility metrics, the XGBoost emerged as the superior model. The top 10 predictive variables identified were malignant arrhythmia, acute kidney injury, Sequential Organ Failure Assessment (SOFA) score, breath sounds, valvular disease, systolic blood pressure, diuretics, blood urea nitrogen, SpO2, and cardiac catheterization. The area under the receiver operating characteristic curve for the XGBoost model was 0.861 [95% confidence interval (CI): 0.836-0.886] with a Kappa value of 0.518 in the internal validation set, and 0.815 (95% CI: 0.799-0.831) with a Kappa value of 0.437 in the external validation set.

CONCLUSION

The XGBoost model exhibited robust discrimination, calibration, and clinical utility, making it a robust tool for risk stratification of prolonged ICU stays in heart failure patients.