Predicting right ventricular failure after lvad using machine learning on mimic-iv
E Vice, P Darko, B Otchere, P Berchie, C Chinnatambi, B Demoss, A Krishnamoorthy, R Singh, R Loungani, K Bauza, S Damle, E Molina, C MartiAbstract
Background
Right ventricular failure (RVF) is a frequent and serious complication following left ventricular assist device (LVAD) implantation, contributing to prolonged intensive care, need for additional mechanical support, and excess mortality. Existing RVF prediction models often rely on invasive haemodynamics or specialised imaging, limiting real-world applicability. There is a need for pragmatic approaches that leverage routinely available clinical data to identify patients at risk for RVF early after LVAD implantation.
Purpose
To characterise the frequency of RVF after LVAD implantation using clinically meaningful proxy definitions and to develop an internally validated prediction model using routinely available early laboratory data.
Methods
We conducted a retrospective cohort study using the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Adult patients undergoing LVAD implantation were identified using ICD-9 and ICD-10 procedure codes. RVF was defined using a composite proxy incorporating post-LVAD extracorporeal membrane oxygenation or right ventricular assist device placement and prolonged inotrope or pulmonary vasodilator use. Baseline laboratory variables from the first 24 hours were extracted. Multivariable logistic regression was developed in a training cohort with internal validation in a held-out test cohort. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC).
Results
The final analytic cohort included 294 LVAD admissions. Mean age was in the mid-60s, and 25.2% of patients were female. Race and ethnicity variables were not consistently available in this extract and were therefore not included in adjusted analyses. RVF occurred in approximately half of patients using the composite proxy definition, driven primarily by prolonged inotrope exposure, while procedural right-sided mechanical support was less frequent. Patients who developed RVF had higher early creatinine, blood urea nitrogen, lactate, and total bilirubin levels, alongside lower sodium, chloride, and bicarbonate concentrations. In multivariable analysis, lower serum chloride and bicarbonate were independently associated with RVF, while increasing age was inversely associated with risk. The model demonstrated moderate discrimination in the training cohort (AUC 0.699) but limited performance in the test cohort (AUC 0.523).
Conclusion
RVF after LVAD implantation is common and often manifests as prolonged pharmacologic support rather than immediate mechanical right-sided failure. Early metabolic and renal laboratory abnormalities are associated with RVF risk, supporting a role for venous congestion and metabolic stress pathways. However, predictive performance using routine laboratory data alone was modest, underscoring the multifactorial nature of RVF. Future models integrating haemodynamics, imaging, and temporal trajectories are likely required for clinically actionable prediction.For image description, please refer to the figure legend and surrounding text.For image description, please refer to the figure legend and surrounding text.