Machine learning to identify attributes that predict patients who leave without being seen in a pediatric emergency
Julia Sarty1, Eleanor A Fitzpatrick2, Majid Taghavi3
1Department of Industrial Engineering, Dalhousie University, Halifax, NS, Canada.
Insights
Machine learning models can predict pediatric patients leaving the Emergency Department (ED) without being seen (LWBS). Key factors include ED patient load, triage hour, and travel time, enabling targeted interventions.
Area of Science:
- Emergency Medicine
- Health Informatics
- Machine Learning Applications
Background:
- Left Without Being Seen (LWBS) is a significant issue in pediatric Emergency Departments (EDs).
- Predictive modeling can help identify at-risk patients for targeted interventions.
- Understanding LWBS drivers is crucial for improving patient flow and care quality.
Purpose of the Study:
- To characterize patients who left without being seen (LWBS) from a Canadian pediatric ED.
- To develop machine learning models to predict LWBS based on key patient attributes.
- To identify the most influential factors associated with LWBS in this population.
Main Methods:
- Analysis of administrative ED data from a Canadian pediatric hospital (April 2017-March 2020).
- Utilized supervised machine learning binary classification algorithms (including XGBoost) to predict LWBS.
- Employed Synthetic Minority Oversampling Technique (SMOTE) for dataset balancing and grid search for hyperparameter tuning.
Main Results:
- The study analyzed 101,266 ED visits, with 5.7% resulting in LWBS.
- The top-performing XGBoost model achieved 95% recall and 87% sensitivity in predicting LWBS.
- Most influential predictors included ED patient load, triage hour, driving distance, length of stay, and patient age.
Conclusions:
- Machine learning models effectively predict LWBS using administrative data in Canadian pediatric EDs.
- Identified five key attributes that significantly influence LWBS predictions.
- Findings support the development of individual patient-level interventions to mitigate LWBS and improve ED efficiency.
Purpose:
To characterize patients who left without being seen (LWBS) from a Canadian pediatric Emergency Department (ED) and create predictive models using machine learning to identify key attributes associated with LWBS.
Methods:
We analyzed administrative ED data from April 1, 2017, to March 31, 2020, from IWK Health ED in Halifax, NS. Variables included: visit disposition; Canadian Triage Acuity Scale (CTAS); triage month, week, day, hour, minute, and day of the week; sex; age; postal code; access to primary care provider; visit payor; referral source; arrival by ambulance; main problem (ICD10); length of stay in minutes; driving distance in minutes; and ED patient load. The data were randomly divided into training (80%) and test datasets (20%). Five supervised machine learning binary classification algorithms were implemented to train models to predict LWBS patients. We balanced the dataset using Synthetic Minority Oversampling Technique (SMOTE) and used grid search for hyperparameter tuning of our models. Model evaluation was made using sensitivity and recall on the test dataset.
Results:
The dataset included 101,266 ED visits where 2009 (2%) records were excluded and 5800 LWBS (5.7%). The highest-performing machine learning model with 16 patient attributes was XGBoost which was able to identify LWBS patients with 95% recall and 87% sensitivity. The most influential attributes in this model were ED patient load, triage hour, driving minutes from home address to ED, length of stay (minutes since triage), and age.
Conclusion:
Our analysis showed that machine learning models can be used on administrative data to predict patients who LWBS in a Canadian pediatric ED. From 16 variables, we identified the five most influential model attributes. System-level interventions to improve patient flow have shown promise for reducing LWBS in some centres. Predicting patients likely to LWBS raises the possibility of individual patient-level interventions to mitigate LWBS.
Related Concept Videos
01:18Steps in Outbreak Investigation
01:24Kaplan-Meier Approach
01:11Residuals and Least-Squares Property
If the observed data point lies above the line, the residual is positive, and the line underestimates the actual data value for y. If the observed data point lies below the line, the residual is negative, and the line overestimates the actual data value for y.
The process of fitting the best-fit...
01:52Regression Toward the Mean
01:10Documentation of Nursing Diagnosis
In some settings, data-driven computerized decision support systems are in place, allowing for more accurate nursing diagnoses. The database within one of these systems includes diagnostic labels defining characteristics, activities, and indicators for nursing. A nurse enters...