All projectsView source
Classification
Length-of-Stay Optimisation Project
Prolonged hospital stays increase costs, reduce bed availability, and can expose patients to hospital-acquired infections. This project modelled factors contributing to extended stays.
Predicted extended stays with 74% accuracy
2025PythonPandasScikit-learnPower BI
Dataset
UCI: Diabetes 130-US Hospitals
Key Questions
- Which admission characteristics best predict extended length of stay (>7 days)?
- Do certain diagnoses or procedure combinations consistently lead to longer stays?
- Can early identification of high-risk patients improve discharge planning efficiency?
Methods
- Feature engineering from admission, diagnostic, and procedural data
- Gradient boosted classifier with early stopping
- SHAP analysis for feature importance and model explainability
- Threshold optimisation for clinical decision support integration
Results
The model predicted extended stays with 74% accuracy. The strongest predictors were number of procedures performed, admission through A&E (vs. elective), number of diagnoses, and patient age. Patients admitted via A&E were 2.3x more likely to exceed 7 days.
Extended Stay Predictors (Feature Importance)
Recommendations
- Flag patients meeting high-risk criteria within 24 hours of admission for early discharge planning
- Assign dedicated discharge coordinators to A&E admissions with 3+ diagnoses
- Implement daily multidisciplinary board rounds for patients approaching day 5 of admission
- Track length of stay by admission source to monitor improvement over time
Limitations
The dataset does not include social factors (housing, care support) that significantly affect discharge readiness. The model focuses on diabetic patients and may not generalise across all specialties without retraining.