In a survey designed to understand patient experience, experts discovered that one of the most often cited factors negatively impacting patient experience was caregiver punctuality and no-shows.

No-shows delay patient care and can cause a major loss of revenue for your home care agency. There are several strategies to reduce no-shows in your organization, and creating a no-show predictive model is one strategy that is cost-saving and essential.

In this article, we give you five keys to creating a useful predictive model that can identify appointments with a likelihood of caregiver no-show.

Key #1: Define the Problem

The goal of this model is to use historical appointment data to predict future no-show events. Here, we define a no-show as either a caregiver:

  1. arriving late for an appointment
    1. Each agency can set a threshold of what is considered “late”, depending on their standards and needs.
  2. cancelling an appointment
    1. Lead time (i.e., how long before a caregiver calls to cancel an appointment) can also be configured. For example, any appointment that is cancelled within 24 hours prior to the start of the appointment is one setting in our scenario.
  3. not showing up at all for the appointment

We can exclude those appointments from our data set where the patient initiated the cancellation.

Key #2: Define the Operational Workflow of Model Output

This step involves using the output of the model’s prediction and integrating it into your workflow to perform actions to mitigate a no-show. Two ways this model’s prediction can be used to mitigate a no-show is via communicating with the caregiver and/or changing scheduling patterns.

Communicating with Caregivers

Sending reminders and connecting with caregivers about any scheduling challenges or updates are the key to ensuring open lines of communication and accountability with your team. By communicating often with caregivers to build strong relationships, you can better use and improve the output from your predictive model to stratify appointments.

Changing Scheduling Patterns

You can use predictive analytics to direct appointments to caregivers who have lower risk of a no-show. The model’s prediction can be presented during the workflow, when an appointment is being scheduled, and give a notification prior to scheduling a visit if a caregiver with a lower probability is available to take the appointment.

Note: Ethical concerns should be reviewed while building feature sets to avoid discrimination against caregiver categories protected by law. This will be discussed further in Step #5.

Key #3: Design the Model

Population Data

Your predictive model should be trained on enough data so that it has localized understanding but not too narrow. What population is chosen for the model will have an impact on no-show patterns and should be considered. The more data that is reflective of your population, the better the model will perform.

Note: If you choose a data time range coinciding with the COVID pandemic onset, this might impact the results of your model’s recommendations.

Training vs Test Data

The model data should be split into 2 categories: Training data and test data. The model learns from training data, which is 80% of the population data chosen for this project. Test data, the remaining 20%, is the data used to check if the predictions produced by the model are reflective of the ground reality.

Feature Selection

Features are defined as variables that will go into the model to help you predict the outcome. You might choose several features that are likely to have an impact on the model’s performance.

Your feature set might be split into external and internal data.

External data can be circumstances, such as weather or traffic, that might impact no-show patterns and can be collected from external sources.

Internal data are circumstances captured inside your scheduling system. This includes such information as appointment day of the week, hour of the day, month, lead time (i.e., hours or days from appointment being scheduled to appointment time), distance from caregiver’s last appointment location (or home) to the location of the next appointment, etc.

The model’s performance might change based on feature selection. Performance should be evaluated using training and test data before deciding on a final set of features to see which features actually impact the model.

Key #4: Understand the Model’s Performance

It is commonly understood that the performance of a model is dependent on the data available to find trends in the underlying structure.

How we understand model performance is through Positive Predictive Value (PPV), or True Positive rate. This means how likely the model prediction will be accurate when classifying a particular appointment to be a no-show.

We want to be correct when labeling an appointment as a no-show to minimize resources on appointments that will not become no-shows.

Key #5: Evaluate Ethical Considerations & Model Impact

Changing the scheduling patterns for a caregiver who has a high risk of no-show can be a controversial practice if the impact of the model discriminates against populations protected by law. You should be proactive in understanding the impact of the model and mitigating this outcome.

Balance between model performance and feature removal

You might consider removing certain demographic features from the model if they present ethical concerns. This step should be weighed as you evaluate model performance with and without those features.

Conclusion

Success lies in better decision making. And better decision making can come from having the data to better predict outcomes. If an organization is unable to accumulate and process its data, it will have little insight into how to use its data to drive better business results. Accumulating data is not enough. You must have a strategy to leverage your data for analytics and better decisions, an ability that can be a differentiator for your organization, helping you improve patient experience and minimize revenue loss. Among other tools, a robust predictive analytics model is an undeniable high value-added asset that home care agencies can use to prevent no-shows and reduce their negative impact.