Article Text
Abstract
Background Emergency Department (ED) overcrowding poses a global challenge, negatively impacting patient safety, satisfaction, and increasing healthcare costs. Long wait times for low-acuity patients are a significant contributor to ED crowding. The aims were to assess the precision of predicting wait times of low-acuity patients (LAWT) using commonly used machine learning (ML) techniques; to ascertain which attributes were good predictors and assess their effect.
Methods The study was conducted at Children’s Health Ireland Crumlin ED, analysing five years of data (January 1st, 2014 - January 1st, 2019). Literature review and localisation of metrics informed input attributes selection to represent crowding. Linear regression (LR), least absolute shrinkage and selection operator (LASSO) regression, random forest (RF) ML techniques were utilised. Root Mean Square Error (RMSE) was utilised to guide accuracy of prediction models. Good attribute coefficients were identified and effect on LAWT established.
Results 190,945 observations were identified. RF was the most accurate model (RMSE 53.32 minutes) compared to LR and LASSO (both RMSE 60.69 min). The top 20 predictors (see table 1) in the RF model included 6 patient factors (e.g., age, number of attendances in the last 1/3 years), 6 temporal-associated attributes (e.g., time of day, month), 5 queue-associated attributes (e.g., wait times of patients ahead in the low-acuity queue) and 3 hospital-associated attributes s (e.g., median time in ED of patients registered in the last two hours, if an EM consultant saw patient within the last 2 hours). Wait time of patient one-ahead in low acuity queue is the top variable.
Conclusion ML models can effectively predict LAWT in a paediatric ED, though the error margin may still be unsatisfactory. Temporal, patient, hospital, and queue attributes all play key roles in predicting LAWT, suggesting that real-time data on ED conditions could provide valuable insights for improving patient flow management.