Building effective predictive models
Predictive models require data. Building, testing and refining these models require data that describes 1) what’s known at the time a prediction needs to be made, and 2) the eventual outcome. For example, to develop a model for heart attack risk presented by patients coming into the ER, we’d need to have data describing patient symptoms when they arrived, and then the subsequent outcome (were they suffering a heart attack or not). The ability to generate data with these characteristics is a critical factor in the success of a predictive modeling application.
Statistical techniques, such as linear regression and neural networks, are then applied to identify predictors and calculate the actual models. After assembling the data, the analysts may find 20 predictive factors that are known for each patient (in our ER example) and assign weights to them using statistical software (e.g., +50 points for abnormally low blood pressure). The statistical software uses algorithms to optimize the model weighting factors, so that the combination produces the most accurate predictions possible with the available data.
The resulting “score,” combining all the factors and their weights, will be an effective risk index that can be used as a decision criterion along with other rules for patient treatment. The score will be not only correlated with cardiac risk, but can be calibrated to have a specific mathematical relationship. This is a crucial advantage; it takes this risk from being an “unknown unknown” to a “known unknown” or, in other words, a calculated risk.
High-value applications for predictive modeling
Most business processes in most organizations have the potential to benefit from predictive modeling. That said, there are certain situations where predictive models can be especially beneficial in delivering a great deal of value:
Processes that require a large number of similar decisions
Where the outcomes have a significant impact, i.e., where there’s a lot at stake in terms of money or lives
Where there’s abundant information in electronic data form available on which to base decisions and measure outcomes
Where it’s possible to insert a model calculation into the actual business process, either to automate decisions or to support human decision makers
This paper discusses how predictive models are built, ideal situations for applying them, calculating their return on investment, key predictive modeling trends and more.