Deep Neural Forecasting in A&E 📈

📊 Poster · 📄 Paper (in progress)

Emergency department (ED) crowding is a leading issue in many EDs and lead to worse patient outcomes. The ability to accurately forecast the occupancy and breach performance in ED a signicant time in advance would be operationally useful, allowing for better allocation of staffing resources, while accurate prediction of admissions would allow for more effective bed management.

In a health system stretched for resources, accurate forecasts would allow us to more effectively use the resources we already have.

In this study, we used the Addenbrooke's Hospital electronic health record system (EHR) to collect minute-by-minute data over a two year period, and used this train a deep neural network to accurately forecast ED occupancy, admissions and breach performance for each hour in the subsequent 24 hour period and evaluated its performance compared to existing techniques.

  • Created a dataset far more extensive and high-resolution than any previous dataset on hospital admissions (minute-by-minute for two years)
  • Deep neural networks are a very good tool for predictive forecasting of ED demand and hospital admissions.
  • Able to pull out non-linear patterns from the previous 96 hours and performs much better than traditional statistical techniques.