Abstract
Emergency Department (ED) overcrowding is a critical global healthcare challenge. This study introduces a transparent data analysis framework using interpretable machine learning (ML) models to predict patient length of stay (LOS) and disposition decisions (DD). By utilizing Lasso regression and innovative feature engineering, the models achieved high predictive accuracy (AUC up to 0.88) while maintaining clinical interpretability, offering a reproducible blueprint for hospital resource management.
Introduction
Overcrowding in Emergency Departments compromises patient safety and operational efficiency. Early prediction of which patients will require admission or a stay exceeding four hours allows for proactive bed management and staffing. However, existing ML solutions often suffer from the "black box" problem—where complex algorithms lack transparency—or a lack of transferability between institutions. This research addresses these gaps by developing a framework that prioritizes interpretability without sacrificing performance.
Methodology and Data Pipeline
The researchers evaluated eight algorithms, ultimately selecting Lasso (Least Absolute Shrinkage and Selection Operator) for the final models. Lasso was chosen because it performs simultaneous variable selection and regularization, resulting in lean models where the influence of each factor is clearly defined by a single coefficient.
Key Innovations in Feature Engineering
To maintain model simplicity while capturing complex clinical realities, the authors implemented several key strategies:
- Target Encoding: Categorical variables were converted into logit values of their associated outcome probabilities, keeping the model structure efficient.
- Addressing Non-Linearity: Standard linear models often fail to capture "U-shaped" relationships, such as blood pressure, where both extreme highs and lows indicate risk. The researchers segmented numeric features into 25 intervals to allow the linear Lasso model to map these non-linear trends accurately.
- Multi-stage Assessment: Predictions were generated at 10, 60, and 120 minutes post-triage to align with the natural progression of clinical workflows.
Dataset and Feature Selection
The study utilized a massive dataset of 297,392 ED visits from a quaternary hospital in Melbourne, Australia, spanning 2019 to 2022. From an initial pool of 170 features in Electronic Patient Records (EPR), the models were refined to the top 21 predictors to ensure they remained manageable for clinical review.
Key Results and Insights
The interpretable Lasso models demonstrated "outstanding performance," proving that transparency does not require a performance penalty:
- Length of Stay (>4 hours): Achieved an Area Under the Curve (AUC) of 0.862 to 0.878.
- Disposition (Admit vs. Discharge): Achieved an AUC of 0.839 to 0.863.
Clinical Predictors
One of the strongest predictors identified was "Order Count" (the number of tests or medications requested). Each unit increase in the processed order count value increased the odds of a long stay by approximately 85%. Interestingly, while "average waiting time" predicted total LOS, it did not significantly influence the clinical decision to admit or discharge a patient, suggesting that clinical judgment remains independent of ED volume pressures.
Conclusion
This research provides a scalable, transparent framework for ED flow prediction. By demonstrating that interpretable models can match the performance of complex ensembles, the study paves the way for greater clinical trust and easier implementation across diverse healthcare settings. Future work will focus on continuous "any-time" predictions and rigorous bias analysis to ensure algorithmic fairness across patient demographics.
Summary updated