5 Common Predictive Modeling Mistakes to Avoid at Your Health System

Big data has become the ”promised land” for healthcare – yet the industry is still learning how to put that valuable data to use, including when marketing to patients.

Most health systems have accumulated massive quantities of patient, provider, and market claims data at this point. However, they often lack robust health analytics to understand the behavior of current patients and populations they’d like to target. 

Fortunately, healthcare-specific marketing technologies are becoming more sophisticated, and most healthcare CRM platforms include built-in predictive analytics. Propensity modeling in particular is a must-have for successful digital marketing teams. It’s a statistical approach that’s used to predict the likelihood that a specific event will occur. Marketers may apply a range of diagnostic, demographic, encounter history, and other variables to predict the likelihood that a patient is, say, a viable candidate for bariatric surgery. If yes, the patient is added to the target audience for a bariatric campaign and receives messages according to rules for that audience. 

Propensity modeling makes the wealth of data available to health systems more actionable. First, propensity models consider a wide array of variables and mathematically condense them. Next, the models derive patterns and relationships from discrete datasets and translate them into key insights. These insights allow marketers to build more strategic, impactful campaigns. 

The process of building propensity models, however, is not simple. Analysts who don’t carefully vet their data sources or who succumb to so-called “overfitting” their models see campaigns fall flat, wasting a great deal of time, money, and other resources.

Here are the five most common predictive modeling mistakes to keep in mind when building propensity models, along with tips for how to avoid them. 

1. Building Models with Incomplete or Inaccurate Data

Thoroughly vetting data before building propensity models is a critical first step. User error, bias, accidental deletion, and improper coding all contribute to a problematic propensity model, and ultimately to problematic campaigns. Healthcare marketers and data analysts must carefully monitor data quality and deal with issues as soon as detected. 

Don’t necessarily assume that technology can find and correct errors. Messy data is best caught intuitively, which is why it’s important to keep both the context and big picture for each model in mind from the beginning. 

Anyone working with the data should have baseline clinical familiarity with it and clearly understand the model’s end use. Data cleaning can be daunting, but tools that help automate univariate and multivariate analysis can expedite the process. 

03 - 509

2. Sticking to One Predictive Model

There’s no such thing as a universal algorithm that constructs the perfect predictive model for all healthcare marketing endeavors. Different algorithms may produce entirely different results from the same datasets. It’s best to test a wide range of algorithms and run your data comparatively to see which ones deduce the most compelling or consistent patterns. Then, look for common points of intersection across each model; these points are likely the most accurate and will indicate that you’re headed in the right direction. By testing multiple algorithms, you’re also less likely to come across false positives in your end results.  

3. Overfitting Models

Overfitting is a common problem for analysts in any field. When creating a predictive model, it can be tempting to adjust your model so it conforms perfectly to your current dataset – to the point that it’s really only describing random error, or noise, as opposed to the legitimate relationship between variables. In other words, an overfitted model contains more parameters than called for by the data.  programming language concept system engineering software development

This problem often occurs when the model has become too complex. Taking a step back to more carefully clean your data is a good remedy. 

Cross-validation also helps, by testing a model across different datasets to ensure it holds real-world predictive power. Additionally, cross-validate them against data that’s already been proven. For example, compare the predicted number of patients with clinical propensity X to the actual number of patients with clinical propensity X. 

4. Believing That Predictive Modeling Replaces Rules-Based Analytics

While predictive modeling reduces the legwork in deciphering healthcare datasets, it doesn’t replace the need for rules-based analytics. All possible conditions and outliers must be built into the model up front to flag problems or outliers. For example, a rule might flag any patient classified under pediatrics who is older than 18. 

While predictive analytics may flag similar outliers by recognizing that such data points do not belong, it won’t do so if certain mistakes frequently recur in your data. If your organization fails to update pediatric patients’ primary care physicians to an adult or family practitioner when they turn 19, there will likely be many improperly classified patients in the system. Thus, a rule is still necessary to flag and correct these outliers. 

5. Failing to Adjust or Alter Predictive Models Over Time

close up group of businessman use stylus pen for explaining dashboard about company's profit monthly on tablet and businesswoman holding report paper for collaborate work at meeting event , business finance conceptAs more data is made available – whether it’s your healthcare CRM, day-to-day EHR influx, or new third-party data – it’s important to refine your predictive models to accommodate changes. You should also test them periodically to determine whether they need “routine maintenance.” 

Again, cross-validation is key to maintaining predictive models: Simply compare your predicted outcomes to actual outcomes and you’ll see whether your model’s accuracy has improved or declined. Usually, the more data you acquire, the more reliable your models become. If this isn’t the case, your models likely need adjustment — or the quality of your incoming data revisited. Regardless, you may need to frequently adjust existing models simply to account for the constant flux of incoming information. 

Final Thoughts

How well you refine your list of prospects largely determines the success of a marketing campaign. You want to hone in on those prospects who will find your message extremely relevant and timely, to the point that it feels personalized. Predictive analytics and propensity modeling give marketers the tools to identify which patients are most likely to have a certain medical need, or which are most likely to respond to a message on a particular channel. 

The abundance of data now available to health systems (and more easily accessible via healthcare CRM) makes propensity modeling even more valuable. The more data used in the model-building process, the more accurate the model tends to be. However, more data from more sources means analysts must be thorough when adjusting algorithms and vetting the data to be used for a given model. 

By paying attention to these common predictive modeling mistakes – and taking steps to avoid them – healthcare marketers craft more effective, data-driven campaigns that drive continued growth for years to come. 

03 - 510



Evariant provides a leading healthcare CRM solution suite designed to help health systems transform the healthcare experience for their consumers, patients, and physicians. Built on the Salesforce Platform, our solutions foster richer consumer/patient engagement and tighter physician alignment. Powered by cutting-edge data and analytics, Evariant enables health systems to effectively communicate care options that increase revenue and market share, while optimizing network utilization. Many of the top health systems have selected Evariant to thrive in today’s hyper-competitive and rapidly changing environment.