Ep 1 - Forecasting patient demand and staffing need with Scott Duncan

Published by Hiro Kawashima on May 14, 2019

Our guest is Scott Duncan, the founder and CTO of Prescience Health. Scott has more than 25 years of experience in the healthcare field and served in leadership roles at the Sachs Group and SG2. He is a mathematician by background and has expertise in developing advanced resource forecasting algorithms.

HIRO: What is the difference between forecasts and predictions?

SCOTT: A Forecast is a mathematical process, usually a projection of the past into the future. For example, in an Emergency Department, if we assume that whatever has historically caused patients to visit the ED will continue for the next few days, then we might get a forecast like: there will be 80 visits to the ED during the DAY shift today.  It’s just a Number. Now, a Prediction adds additional context to make a forecast more useful. In our example, a good Prediction might be: it is 70% likely there will be between 75 and 85 Visits to the ED during the DAY shift today, so there is a need for 4 nurses, and you only have 3 nurses on the schedule, so you should call in another nurse. An even better prediction might say that the severity of those patients is expected to be higher than normal, so you should try to call in a Senior nurse, or maybe even two more nurses.

HIRO: Why is it so difficult to forecast patient demand?

SCOTT: Like Real Estate, Healthcare is extremely LOCAL.  No two hospitals are the same, and no two departments within a hospital are the same.  The best forecasting approach for an ICU at an Academic Medical Center is not going to be the same as for an ICU at a Community Hospital, and certainly not the same as the NICU, or Telemetry, or Women’s Services: other departments within the Medical Center.  Adding to the difficulty is the influence on Patient Patterns of Holidays, Weather, Pollen Counts, School Calendar, etc. And, of course, a single event, like a Bus Crash or a Music Festival or a Winter Storm, can throw off the numbers for several days. Any useful approach to forecasting patient volumes needs to be flexible and reactive, and customizable to each individual department.  In other words, designed specifically for how Healthcare works.

HIRO: What are some methods you use to forecast patient demand?

SCOTT: At Prescience Health, we employ a forecasting approach called Box-Jenkins, which is best described as a model-competition approach.  For each department, we try out multiple forecasting models, some dynamic, and some more static, and then judge each one on how well it would have predicted the recent history for that department.  If the patterns for a department change drastically, maybe due to physician admitting changes, or market competition changes, etc, the Box-Jenkins approach will notice and switch to a new model that is more appropriate for the new patterns. We then layer on top of that base forecast several proprietary techniques that improve its usefulness in Healthcare specifically. Then we combine the forecast with the staffing and scheduling guidelines of each department to generate the staffing levels by SHIFT that will provide for that department the best BALANCE between COVERAGE and COSTS, as defined by each department manager individually.

HIRO: Besides patient census, what other demand drivers can you forecast?

SCOTT: Our StaffRight engine can forecast anything that directly impacts patient care requirements and has historical data.  For example, in the Inpatient setting we often forecast not only Census but also Acuity and ADT Activities (Admission, Discharge, Transfers). This allows Inpatient department managers to go beyond the traditional Staffing Grid to more accurately translate patient volumes into staffing levels needed.  For example, in an ICU, the staffing guidelines can be set to say that patients of acuity level 4 require 30 minutes of nursing care each hour, but patients of acuity level 5 require 60 minutes of nursing care each hour, and an additional 20 minutes of nursing care is required for each Admission, Discharge or Transfer activity. Similarly, in an Emergency Department, we may be asked to forecast Visits, AND Census AND Severity.  For Women’s Services, in IntraPartum, we may forecast Labor with & without complications or Cesareans, and in PostPartum we may forecast different levels of recovery needs for the mothers, and in the Nursery we may forecast Normal Newborns versus continuing care, intermediate care, or intensive care newborns. There is tremendous flexibility in what demand drivers we can apply to determine each individual department’s staffing and scheduling needs.

HIRO: What factors influence the accuracy of the forecasts?

SCOTT: Higher volume departments with lower variability and a fairly steady ALOS are easier to forecast. For example, if your Telemetry unit has 20 beds and is ‘always full’, then your forecast for next Tuesday is 20.  Units with lower volumes, and higher variability, such as NICU or Labor & Delivery are much, much harder to forecast, and therefore much harder to staff efficiently. To help manage these more difficult departments, our forecasts are continually updated and improved throughout the day, and our StaffRight product has Active Alerting to notify managers and staff of any important changes to the staffing needs. StaffRight also provides tools to ‘see’ the staffing needs forecasts across several units simultaneously, to help managers balance staffing over multiple related units.  Again, every department is different, and any product that claims to guarantee a certain level of forecasting accuracy is not being honest.

HIRO: What are your thoughts on machine learning/AI and should it be incorporated into patient demand forecasting?

SCOTT: The key word here is ‘volume’.  Machine learning/AI is not currently very helpful in forecasting how many patients you will have next Tuesday, but there is much promise that someday soon it will help with determining the staffing needs for the patients already in your hospital, by predicting likely clinical pathways and procedures for individual patients, and then matching that up to the capabilities of staff to determine the best staffing to provide the best care, given the current (and near future) case mix in each department. In other words, machine learning/AI can potentially supplement patient demand forecasting to further optimize the patient care your staff can provide to your patients.

Next post: Ep 2 - Using forecasting to optimize unit coverage with Chris Looby

Take the guesswork out of staffing

Schedule a demo and learn more about how predictive analytics can help you staff better