Patient Volume Forecasting in Hospitals and Care Providers with IBM PA

Accurate forecasting of patient volumes is critical to help hospitals and care providers optimize allocation of resources. This Solution Brief outlines an application to support effective forecasting using IBM Planning Analytics (IBM PA). The document is broken down into the following three areas:

1. Automated Patient Volume Tracking, Reporting and Analysis
2. Patient Volume Forecasting using Statistical Models and AI
3. Scenario Planning and What-If Analysis for Resource Allocation

Automated Patient Volume Tracking, Reporting and Analysis

The initial step is to integrate the IBM PA platform with internal systems to automate reporting and analytics. This provides real time insight into patient volume metrics across any dimensions required. Information is aggregated by hospital, department but also by region or any other custom criteria to provide automated reporting and analysis of data.

Access to information in real time is key to help analyze trends and measure efficiency of schedules to handle patient volumes.

In a fully integrated application, IBM Planning Analytics is directly connected to source systems through embedded APIs and connectors. Data is stored in the system at a granular level with several attributes as needed for analysis. Information can be tracked on a daily or even hourly basis by department, location or other variables based on each individual hospital system structures. Data can be updated daily, weekly, monthly, or even intra-day and on demand as needed.

In the chart below, you can see a clear weekly pattern of Emergency Department visits by day for a specific location.

Some of the key benefits and outcomes once data is loaded are the following:

• Real Time Availability - Once loaded, all data in IBM PA is aggregated in real time and available for reporting. All calculations are performed in real time, including ratios and KPIs as defined by the model
• Variance Analysis - Actuals are tracked separately from budget and forecast. Users can analyze each scenario separately and dive into variances vs budget and re-forecast periodically (monthly, weekly, daily) to adjust to the most recent information.
• Automated Reporting - Reports are automatically generated for the analyst and distributed to the respective stakeholders.
• Free form analytics – unlimited dimensions and attributes to get insight into the data, free-form pivot-table style analysis of data
• Dashboarding and visualizations – user-driven custom dashboards to present results, metrics and KPIs

Example dashboard available in the system:

Patient Volume Forecasting Using Statistical Models and AI

Variables Affecting Patient Volumes

Patient volumes at any hospital are driven by a number of complex variables, which may differ by location, time of the year, department and other. To effectively estimate future volumes with any level of precision, hospitals need to leverage a number of internal and external data sources.

• Internal Data – historical volume patterns can be used as a base indication of expected volume. In the charts above, we can observe a very clear and predictable pattern of volume based on the day of the week as well as the time during the day.
• External Factors - This pattern can be enhanced by external data that can influence the volumes and create variations based on specific events. These events will vary by department, location etc, some examples that we observed / used include the following:
  • Inpatient encounter historical data within the MSA/market (overlay hospital-specific volumes with local and regional data)
  • CDC Flu Sight- Predicts flu projections taking historical data with trends to reflect potential volume hitting the ED in different markets.
  • Impact of Snow events – such as >5 inches and extreme cold for more than 2 days. In some markets, hospitals observed the 8-10 hours leading up to a snow event resulted in roughly half the volume and the first 6-8 hours after snow ended the volume was increased around 30%.
  • Changes in public transportation schedules, especially bus schedules, can have a significant impact on the hourly arrivals. 

Statistical Patient Volume Forecasting

In IBM Planning Analytics, these variables can be incorporated seamlessly into the forecast. They are used as a basis for multivariate statistical forecasting, which is available natively inside the system. The algorithm uses historical patterns based in internal patient volume data for each particular location, department etc. The internal data is then enriched with external factors that can be sourced from the market though and automated interface and factored into the estimate.

The statistical forecast uses the ARIMA, ARIMAX or VAR models depending on each specific data set – the methodology is selected automatically by the system based on the outcomes statistical model metrics. Users can select parameters, such as confidence intervals, but also select time periods and remove outliers to refine the forecast methodology and arrive at desired outcomes.

The estimates are often required by hour as they need to be factored into specific shifts and schedules for different practitioners.

The statistical forecast can be run multiple times in parallel using different assumptions and variables. The system will display the model parameters that describe the statistical model and will provide transparency as to forecast accuracy (picture below). Users can compare the outcomes under different assumptions side-by-side and chose the one that gives them the most accurate result.

Scenario Planning and What-If Analysis for Resource Allocation

Once the patient volume forecast is in place, the next question is how many patients will there be in a specific department requiring care at any point in time. This is a function of a number of variables, in our Emergency Department example we focused on three key ones:

• Patient arrival volume – this was determined in the previous steps described above
• Average Length of Stay (ALOS) – usually expressed in hours, it determines how long an average patients stays in the ED receiving treatment
• Case Mix Index (CMI) - a measure that represents average clinical complexity or severity of illness. It determines the resource intensity requirements for each specific patient

Using these variables, we can model the number of patients that will require care at any point in time. As we can see the figure below, the number of patients in the ED grows as more people arrive in the morning and accumulate receiving care and then trail off in the evening.

The system allows unlimited number of scenarios that can be analyzed in parallel using different assumptions as part of a what-if analysis. These can be managed concurrently and compared side by side to determine best outcome. Multiple users and departments can collaborate using different assumptions and variables. The model changes in real time based on changes in patient volumes as well as assumptions. 

In the final step, users can use this informaiton to allocate resources and practitioners to individual shifts as needed in a way that optimizes resource utilization. This can be done in IBM PA also or the information here can be exported and used in a separate scheduling system if one is in place.

If you would like to discuss this in more detail or see a demo of the application please contact us at solutions@acgi.com.