A sequential machine learning framework for ICU and hospital length of stay prediction from admission-time data

Objectives: Accurate prediction of hospital length of stay (LOS) is essential for effective patient flow management, resource allocation, and improved clinical outcomes, particularly in resource-intensive settings such as intensive care units (ICUs). Existing LOS prediction models often rely on post-admission data or focus on homogeneous patient populations, which limits their utility for early operational planning. This study aims to develop a sequential machine-learning framework to predict total hospital LOS using admission-time information. The framework predicts ICU admission and ICU duration as intermediate outcomes, which are then incorporated as features to improve total LOS prediction while also providing standalone operational value for early resource planning. Methods: We propose a sequential prediction framework that decomposes LOS prediction into staged tasks, enabling early and interpretable clinical and operational decision support. The framework was validated using a large dataset consisting of 1,104,203 patient admissions from 52 Maryland hospitals, including 20 admission diagnoses and 29 comorbidities. Deep learning and classical machine learning models were employed for both classification and regression tasks, and performance was compared with traditional single-stage LOS prediction approaches. Results: The XGBoost classifier predicted ICU admission with an accuracy of 0.73, sensitivity of 0.76, and specificity of 0.72. Incorporating ICU-specific predictions improved total hospital LOS prediction performance. For LOS classification, the proposed framework achieved an F1 score of 0.701 (a 5.1% relative improvement) and a macro AUC-ROC of 0.76 (a 1.9% relative improvement) compared with single-stage models. Regression analysis resulted in a mean absolute error of 2.62 days for LOS prediction. Conclusion: The proposed sequential, ICU-informed machine-learning framework enhances the accuracy and interpretability of total hospital LOS prediction using admission-time data. By explicitly modeling ICU admission and ICU LOS as intermediate outcomes, the framework provides actionable insights to support early patient flow management, resource utilization, and operational planning within hospital systems.