An ensemble learning approach for predicting hospital stay in transplant patients

The rising incidence of heart and lung failure has increased the demand for effective transplant management strategies. Predicting Hospital Length of Stay (HLOS) is essential for reducing cost variability, optimizing resource utilization, and supporting patient recovery. This study uses data from the United Network for Organ Sharing (UNOS) to develop and validate an Ensemble Meta Stacked (EMS) model for predicting hospitalization duration after heart and lung transplantation. Expert-informed feature engineering incorporates donor and recipient compatibility measures, while a hybrid two-stage feature selection process combines expert evaluation with the Boruta algorithm to identify key predictors across demographic, clinical, behavioral, and geographical domains. Twelve predictive models are developed, including five base learners for each organ type and an EMS model that integrates their outputs through a Random Forest (RF) meta learner. Among the base learners, RF achieves the highest accuracy, but the EMS consistently outperforms all individual models. Sensitivity analysis confirms the robustness of model performance under different feature sources and scaling procedures, while paired statistical tests confirm that the improvement in predictive accuracy of EMS compared to the base learners is not due to random variation. The study also links predictive metrics to stakeholder priorities: policymakers and payers benefit from stable forecasts that control financial variability, hospital administrators rely on consistent prediction accuracy for capacity planning and resource allocation, and clinicians depend on bias-related metrics to guide safer discharge decisions. The EMS framework advances data-driven management in transplantation, supporting more efficient, equitable, and clinically responsible care. • Predict hospital stay durations after heart and lung transplants using integrated machine learning techniques. • Identify critical predictors from demographic, behavioral, and clinical factors with advanced feature selection methods. • Build a stacked ensemble model using five distinct machine learning techniques and improve prediction accuracy. • Demonstrate how model architecture affects the risk of overprediction and underprediction in clinical outcomes. • Support healthcare planning with data-driven predictions that enhance patient care and manage treatment costs.