Novel Machine Learning Approach to Predict and Personalize Length of Stay for Patients Admitted with Syncope from the Emergency Department
Gradient-boosted decision trees trained on the National Emergency Department Sample predict whether syncope admissions will stay more than 48 hours, supporting disposition planning.
Key Findings
- Predictive lift. The best model reached an AUC of 0.88 for the ≤48-hour versus >48-hour task, outperforming logistic regression baselines.
- Determinants of prolonged stay. Age, injury severity codes, cardiac comorbidities, and hospital resource markers carried the highest SHAP importance scores.
- Transparent explanations. Global and encounter-level SHAP plots enabled clinicians to review why the algorithm recommended extended monitoring.
Introduction
Emergency clinicians often admit syncope patients for observation because risk stratification tools do not reliably indicate who can safely leave within two days. A data-driven length-of-stay classifier offers an evidence-based alternative to subjective judgement.
Methods
Nationwide Emergency Department Sample records (2016–2019) for adult syncope encounters were filtered, cleaned, and engineered into clinical, demographic, and utilization features. Gradient boosted models (XGBoost and LightGBM) were tuned via cross-validation to discriminate ≤48-hour stays from longer admissions, with held-out evaluation on a 20% test cohort.
Results
At the operationally relevant 48-hour threshold the classifier balanced sensitivity and specificity around 0.80, providing calibrated stay probabilities that aligned with observed event rates across deciles.
Discussion
Because the model was trained on a national dataset, local recalibration will still be required, but the workflow demonstrates that routinely collected administrative and clinical signals can inform observation decisions without bedside data entry.
Clinical Implications
Embedding the predictions into emergency dashboards could prioritise telemetry beds for high-risk patients while expediting safe discharges, reducing unnecessary admissions and costs.
Conclusion
Machine learning-derived length-of-stay predictions translate syncope big data into actionable triage guidance.
Future Directions
Prospective trials should combine administrative signals with streaming vital signs, incorporate shared decision tools, and test whether the model reduces crowding or readmissions when deployed.