Using Machine Learning to Identify Modifiable Hospital Throughput Drivers of Prolonged Length of Stay

Prolonged hospital length of stay (LOS) is a major driver of hospital crowding and emergency department (ED) boarding, yet the clinical processes contributing to extended stays remain poorly characterized. The objective of this study was to use an artificial intelligence-based data-driven framework, grounded in stakeholder-informed clinical process mapping, to identify modifiable drivers of prolonged hospital LOS.

This contemporary cross-sectional study included adult hospitalizations from the ED to a large urban academic hospital (Yale New Haven Hospital) in 2023, using the top 30 diagnosis-related groups with the greatest observed-to-expected (O/E) variation. Patients with elective or direct admissions, in-hospital death, or those who left against medical advice were excluded. Prolonged LOS was defined as O/E > 1. Structured electronic health record features and data sources were matched with a stakeholder-informed clinical process map (input-throughput-output) and used to train an XGBoost model to predict prolonged LOS. Model performance was evaluated by receiver operating characteristic area under the curve (ROC-AUC), precision-recall AUC, sensitivity, specificity, and SHapley Additive exPlanations (SHAP).

Among 12,771 admissions, the mean age was 67.1 years, with 51.8% females, 47.7% were covered by Medicare or Medicaid, 22.1% were admitted to ICU, and 43.7% had a prolonged LOS. SHAP analysis revealed that throughput-related process factors such as the number of consults, number of imaging studies, and delays in consult and imaging turnaround were the strongest contributors to prolonged LOS. These exceeded the predictive influence of input (admission) factors (e.g., ED volume) and output (discharge) delays (e.g., discharge readiness status). Ultrasound and delays in consults emerged as high-impact, actionable drivers of inefficiency. The XGBoost model demonstrated strong performance (ROC-AUC 0.89, PR-AUC 0.74, specificity 0.90, sensitivity 0.54).

Combining a holistic process mapped lens with a high-performing machine learning model identified modifiable throughput bottlenecks, particularly related to consults and imaging, as key drivers of prolonged hospital stays. Targeted interventions focused on streamlining these processes may reduce LOS, enhance inpatient capacity, and alleviate inpatient boarding in the ED.