Healthcare analytics · Production model
Looking ahead.
Understanding uncertainty.
Operational wait-time forecasting for the Emergency Department, with careful attention to patient-level validation and the consequences of underprediction.
The question
Emergency Department wait times change with patient demand, acuity, and operating conditions. The analytical task was to translate those changing conditions into useful operational predictions.
The approach
I built and deployed a forecasting model using operational and temporal features. These included patient volumes, acuity, ambulance arrivals, admitted-patient counts, recent wait times, time of day, day of week, and holidays.
I evaluated Extra Trees, XGBoost, LightGBM, neural networks, and quantile approaches. Dynamic P80 calibration targeted approximately 80% patient-level coverage; this describes the calibration target rather than a claimed achieved result.
How I evaluated it
The evaluation considered MAE, RMSE, R-squared, bias, pinball loss, P80 coverage, and underprediction rate. Looking at several measures helped assess both typical errors and how the model handled uncertainty.
Why it matters
Operational forecasting needs more than a single accuracy score. Understanding where predictions miss, whether they are biased, and how often they underpredict is part of making model outputs useful.
Public case-study scope
This is a high-level summary of professional work. No patient records, private dashboards, internal systems, or confidential model results are shown.
What matters
A useful forecast needs a clear account of its uncertainty and its limitations.
Keep exploring