A two-stage framework for hospital-wide length of stay prediction using real-world inpatient data: prolonged-stay risk stratification and short-stay estimation

Hospital-wide length of stay (HWLOS) prediction is challenging because inpatient LOS is right-skewed and clinically heterogeneous, while prolonged-stay risk stratification and day-level estimation serve different operational needs. We developed and internally validated a two-stage HWLOS framework using real-world inpatient data. We analyzed 209,088 admissions from 104,245 patients in acute general beds at a medical center in Taiwan (2022–2024) using patient-grouped partitions. Stage 1 used calibrated XGBoost to classify prolonged stays (> 20 days); Stage 2 used a deep neural network (DNN) to estimate day-level LOS among short stays, with linear regression, random forest, and XGBoost as comparators. Single-stage comparisons, predicted routing, SHapley Additive exPlanations (SHAP), and sensitivity analyses of prolonged-stay cutoffs and information-restricted model specifications were evaluated. Stage 1 achieved an AUROC of 0.9479, 80.21% sensitivity, and 33.33% positive predictive value at the validation-selected threshold. On oracle-defined short stays, the DNN had the lowest mean absolute error (1.4609 days) and highest proportions within ± 1 and ± 2 days, whereas XGBoost had the lowest root mean squared error (2.4289) and highest R ² (0.6231); no model was uniformly best. Predicted routing sent 88.11% of admissions to Stage 2; on the same 37,199 routed admissions, there was no clear evidence of an MAE difference between two-stage and single-stage DNNs. Across prolonged-stay cutoffs of 14, 20, and 30 days, sensitivity remained near 80%, while alert burden increased and the Stage 2 short-stay population expanded. SHAP identified admission source, number of procedures, and physician-predicted LOS category as leading predictors. When discharge-finalized predictors were excluded, Stage 1 AUROC decreased from 0.9479 to 0.8555, and Stage 2 MAE increased from 1.4609 to 1.7400 days. The two-stage framework distinguishes complementary prediction tasks: prolonged-stay risk stratification to inform discharge-planning and case-management review, and day-level LOS estimation for predicted short stays to support anticipated bed turnover and capacity planning. Evaluation under predicted routing makes the associated trade-offs explicit. Prospective evaluation using predictor values available at the intended prediction time and multicenter validation are needed before clinical implementation.