Hybrid GAM–XGB Model for Predicting Hospital Length of Stay

Due to the rise in both chronic illnesses and longevity, healthcare systems are now compelled to manage their resources more strategically. In this context, Length of Stay (LOS) stands out as a key performance indicator for measuring the efficiency of healthcare services. This study analyzed the effects of homogram findings and demographic data on LOS for 6560 patients admitted to the cardiology clinic with a diagnosis of decompensated heart failure. To overcome the limitations of existing statistical and machine learning models, a new hybrid model combining Generalized Additive Models (GAM) and Xtreme Gradient Boosting (XGB) was introduced. The transfer of partial effects from GAM into the XGB model, utilizing them as new features, was instrumental in achieving a significant improvement in the model's overall prediction and generalization performance. The results show that the proposed model offers higher performance and clinical interpretability in LOS prediction compared to other models. The proposed method is poised to significantly advance healthcare planning by mitigating critical shortcomings in existing LOS prediction tools and establishing a robust framework for advanced clinical decision support systems.