Improvement of your deployment operations requires that you understand where your services will be needed and how to get the available units into the most suitable positions. Then, once you are prepared to respond, it is also critical that only the most appropriate assignments are made for each request to preserve your ability to respond to the next call as well. Traditionally call assignment was a simple “closest unit” consideration with all your resources being equal. That task has now become increasingly complex with a recognition of a growing diversity in call acuity and the increasingly common tiered capabilities of your immediately available resources. This second step of appropriate dispatching toward operational efficiency will be the subject of a future blog post to focus this article solely on demand forecasting.
A common practice for emergency services that have grown beyond a single central depot has been a simple distribution of their resources geographically in the hopes of being able to serve anyone at any time. Without an abundance of crews, this is not typically a successful strategy since neither population nor risk are ever uniform. To make matters even worse, most agencies are experiencing an increase in their volume of demand while also facing some of the most serious challenges in decades to simply maintain staffing levels. Emergency medical services across the country are reporting employee turn-over rates of around a quarter of their staff annually. This trend suggests that there are not only fewer providers per call but less experience on each transport as well. Dispatch centers across the nation also face challenges with an average of 20% of their staff positions routinely left unfilled during the past few years. These difficulties underscore the importance of making good decisions quickly.
Figure 1
Disproportionate access to services is difficult to resolve with fixed stations and often results in increased service available outside of the intended district. To adequately populate this geographic coverage model requires an excessive workforce.
So, are you accurately forecasting demand to help improve your operations?
Is your current demand forecasting process recognizing trends throughout the day and week to allow for effective decision making in response to any predicted demand patterns? Without some certainty in your predictive capabilities, it is impossible to effectively trust the recommendations of any decision automation. A potential lack of credible information makes the choices of unit movement more difficult at the same time they are becoming even more critical to the agency. And a lack of credibility also encourages the freelancing of decisions outside of the control of your administration.
Seasonal variation
When reviewing your own annual call history, you should notice the seasonal variation that distinguishes not only the volume of calls within or between school calendars, but the very nature of the calls themselves tend to follow a pattern. During the summer months, personal schedules tend to be increasingly variable with more adventurous outside activities repeatedly lead to more traumatic events. Once school is in session, most families have less-flexible schedules and the shorter, cooler days often make individuals more vulnerable to acute medical conditions.
Even shorter temporal variation
On a shorter scale of time, differences are also recognizable by day of the week or even hour of the day. Higher call volumes typically occur toward the start and end of the traditional work week. The early morning hours of these business days also exhibit patterns found with early waking habits and the increased vehicle traffic and population movement. The pattern repeats itself later in the afternoon, but the locations of people are quite different than in the morning. The unique business hours and personal behaviors on the weekend also make these days unique from the rest of the week. An unequal distribution of people throughout space and time leaves discernable patterns in the location of requests as well.
Figure 2
Sample data demonstrating ALS (green) and BLS (blue) call volume comparisons by hour-of-the-day and day-of-the-week. Notice the similarity in daily patterns although total volume (represented by 90th percentile) is unique.
To create a useful model that honors all these variations, the operational period to be described in a forecast must generally be shortened while simultaneously extending the pool of similar examples to achieve the required statistical precision. The more similar the forecast of demand is to the current moment in time, the more useful it will be in guiding effective decisions. If the intention is to describe demand during the next hour or two, the historic records queried should reflect a comparable timeframe.
Fortunately, your call history is proven to contain many useful clues about the future. It is not merely a matter of extrapolating a progression of time or an assumption that the same request will come from the same caller again. The real-world is complex, yet we all tend to live, work, and associate with individuals that are more like us than the overall population. As a result, each previous request is an indicator of the types of requests likely from our unique population cohorts. The successful technique is in the allocation of the right populations within the right timeframes to sufficiently forecast the future demand. This is accomplished through the way incident records are selected in a dynamic query to represent a time-based forecast and even more importantly, how those results will be spatially aggregated.
Through Demand Monitor, MARVLIS users can not only update forecast parameters based on their local knowledge, but they can monitor both the accuracy and precision of each dynamic forecast. Using a default configuration, most services should find that approximately 80% of the actual calls received are in an identified hotspot recognized by a current demand forecast. With some effort, that average can often be raised to over 90% of future requests are correctly forecast by the hotspot zones. Simply raising accuracy, however, could be easily accomplished if precision is not considered. By including the whole jurisdiction within a hotspot, an accuracy of 100% would be the result. While technically valid, this type of forecast would provide absolutely no assistance in pre-positioning responders to improve outcomes. The forecast area must be maintained as small as possible while increasing the predictive capabilities of a demand query. Currently, this is recognized by comparing incoming requests over time to the effective forecast when each call was received.
Demand Monitor allows analysts to define multiple query strategies for simultaneous execution and evaluation. If each of these queries is validated against reality, the distinct forecasts can be quantitatively compared and improved over time. The result is a continuous quality improvement that requires some regular review to maintain.
Best practices in Demand Monitor
A recommended best practice for modeling demand is reviewing and modifying the demand queries at least twice a year to coincide roughly with the school calendar. It is not necessary to be precise in modelling academic dates, it is the mindset of schedule regularity that is driving the demand pattern. Jack Stout, the father of the System Status Management concept, suggested using a floating 20-week period based on the size of the spreadsheets he used, but this often crosses the known seasonal variations discussed earlier. To minimize the impact of influence from outside the current season, the number of weeks can be shortened. Reviewing only 5 weeks before and after the current forecast date cuts that total number of weeks in half. It is possible to maintain the number of records of the longer period by including the same weeks from a previous year to mitigate the reduction of number of samples while maintaining seasonality. However, the addition of too many years may have a detrimental affect by increasing the influence of older neighborhoods since newer subdivisions would have less representation across the years. Experience suggests looking back no further than 2 previous years in most circumstances. For most agencies, that keeps the records reviewed within the post-pandemic experience as well.
Another successful strategy to control for the temporal pattern can be to query a fixed seasonal timeframe rather than a floating period. If you want to model the school year, setting fixed dates of mid to late August through mid-May will clearly eliminate the effect of any summer dates. A downside to this method would be the necessity of changing the query period once school begins or ends for the year. Automating the model of both strategies simultaneously can allow for each query option to be graded separately to discover the best alternative for your jurisdiction.
It is difficult to argue against modeling each day of the week individually, but when it comes to the finer segmentations of the day, there is legitimate debate. Again, Jack Stout recommended modeling each hour of each day for a total of 168 unique timeframes of the week. Part of his justification is the average busy time of a unit being about an hour and to simplify the calculation of a Unit Hour Utilization (UHU). Demand Monitor is typically automated to execute every 5 to 10 minutes to minimize the amount of change between each forecast while allowing the predictions to subtly adjust more frequently. It is also common for ambulances to be busy longer than an hour in our post-pandemic world.
Once a query definition is set, it can be tested in Demand Monitor to see how many records it will return. Ideally, the number of records for any sample query should be measured in the hundreds, but less than a thousand. If you need to adjust your parameters, altering the number of years will have the greatest impact followed by the number of weeks and finally the number of minutes which will have the smallest influence.
The experts at BCS have decades of experience bringing real-time analytics to the real-world. If you require any assistance in customizing your Demand Monitor queries, please contact your support representative.
High Performance EMS 