부산대학교 도서관

Introduction: The Seattle Heart Failure Model (SHFM) is a validated, widely-used risk model to predict survival in patients with heart failure (HF). The SHFM was developed with a multivariate Cox model, which may not capture complex, non-linear relationships between the predictors and the outcome. We hypothesized that a machine learning model could capture such relationships and therefore improve prediction accuracy using the same variables as the SHFM. Methods: A total of 79,958 echocardiograms were identified from 20,529 patients with heart failure in a large health system (Geisinger). We predicted all-cause mortality for 10 survival durations after echocardiography: from 6 to 60 months at 6-month intervals. The 14 variables (see Figure caption) utilized by the SHFM were used as input variables for 2 machine learning classifiers: linear logistic regression and non-linear gradient boosting trees (GBT). The mean area under the curve (AUC), across 10 cross-validation folds, was obtained for each survival duration and compared between the machine learning models and the SHFM with paired t-tests. Results: The median follow-up time was 6.2 years (range: 2.4 - 11.4). The number of echocardiograms used for each of the 10 survival duration models is shown in the Figure. The SHFM yielded AUCs of 0.72-0.74 to predict all-cause mortality. Both machine learning models significantly outperformed the SHFM for all survival durations (all p<0.001, Figure). The best performing GBT models (all AUC>0.80) revealed that, of the variables used in the SHFM, systolic blood pressure and left ventricular ejection fraction were the two most important predictors of survival. Conclusions: Machine learning models significantly outperform the SHFM in predicting all-cause mortality in patients with HF across multiple survival durations up to 5 years after echocardiography. Machine learning may therefore provide a more powerful tool to risk stratify patients with heart failure. [ABSTRACT FROM AUTHOR]