Machine learning prediction of right heart failure after left ventricular assist device implantation with individual risk factor visualization.

BACKGROUND

Right ventricular failure (RVF) is a major adverse event after left ventricular assist device (LVAD) implantation, but its complex mechanisms make accurate prediction challenging. While supervised machine learning (ML) can predict outcomes from complex models, understanding the conclusions is often difficult. This study aimed to develop an interpretable supervised ML model to assess patients' RVF risk and identify key risk factors.

METHODS

We analyzed 327 patients from 482 consecutive LVAD implantations performed between June 2010 and January 2024 who underwent preoperative right heart catheterization and echocardiography. RVF was defined as requiring a right ventricular assist device or ≥14 days of postoperative inotropic support. Preoperative clinical and hemodynamic variables were used for risk prediction, and Shapley additive explanation (SHAP) values were used to assess individual risk factors.

RESULTS

Thirteen features were selected: sex, age, nonischemic cardiomyopathy, body surface area, aspartate aminotransferase, blood urea nitrogen, left ventricular end-diastolic dimension, left ventricular ejection fraction, cardiac output, right ventricular stroke work index, central venous pressure (CVP)/pulmonary capillary wedge pressure (PCWP), pulmonary pulsatility index (PAPi), and Interagency Registry for Mechanically Assisted Circulatory Support (INTERMACS) profile. Ensemble learning was the most reliable algorithm, achieving an area under the curve of 0.888 (95% confidence interval [CI], 0.756-0.975) and an accuracy of 0.879 (95% CI, 0.788-0.940). A high-risk group (RVF occurrence rate >20%) had 1-year and 3-year survival rates of 80.7% and 75.3%, respectively, showing a tendency toward worse long-term outcomes compared to the low-risk group (93.2% vs 90.5%; P = .09). SHAP analysis revealed individual patient risk factors and their contributions, most frequently highlighting CVP/PCWP, PAPi, and INTERMACS profile.

CONCLUSIONS

Supervised ML enables accurate prediction of RVF after LVAD implantation. SHAP values identify individual risk factors, potentially optimizing preoperative conditions.