Operational Data Challenge

Operational data challenge

We would like to invite all interested in machine learning and operational research to explore our operational dataset. You can download it from here:

This data, describing patient waits in a hospital, was built for the following challenges (see our recent publication here: https://rdcu.be/b4ffC):

  • Predict patient wait times from the other operational features, as accurately as possible
  • Identify the smallest subset of features, sufficient for accurate wait time prediction across all four facilities in the dataset. The smallest model MAE should be at most 1-2 minutes worse compared to the best full model MAE
  • Invent (engineer) new features significantly improving model prediction quality

Overall, reducing MAE by more than 70% in comparison with the simple intercept model (predicting wait from its overall average) would be a significant step forward.

Data description

Our dataset represents operational features, captured in four different hospital facilities, processing walk-in (F4) or scheduled (F1, F2, F3) patients. Approximately 600 to 1000 days of full  patient flow data were extracted from each facility. The target variable to be predicted is Wait. The other variables represent different features of the patient flow, captured at the time of patient arrivals and departure (one line per each patient visit event). Data Excel file contains the exact definitions of all features used (see Contents sheet).

This data was extracted from the real hospital information system. Therefore it was anonymized and aggregated to remove any confidential information. In particular, x_ArrivalDTTM, x_ScheduledDTTM, and x_BeginDTTM timestamps were anonymized, and set to the fictitious future dates (to make you aware of this modification). Only their relative timing was preserved. As a result, these timestamps can be used only for sorting, to reflect the correct order of feature events, but nothing else.

Code

We include our Matlab code used to process this data here – feel free to download and experiment.

%d bloggers like this: