Feature Interaction in Medical Science (and beyond)

In data science, we often focus on building models that can predict outcomes based on individual features or variables, such as age, income, or blood pressure.

But what happens when two or more features interact with each other, affecting the outcome in ways that go beyond their individual contributions? This is where feature interaction comes into play, and understanding its importance can significantly improve the performance of your models.

Let’s take the example of the QRISK model from this medical study.

This model was developed to predict the risk of cardiovascular disease (CVD). Early versions of the model used individual features such as age, smoking status, and cholesterol levels. While useful, these predictions didn’t fully capture the complexity of real-world interactions between risk factors.

The newest version, QRISK3, addressed this by inclusion of feature interactions. Feature interaction refers to how two or more variables in a model affect the predicted outcome in combination, rather than individually.

In QRISK3, interactions between certain variables — like ethnicity and diabetes or age and systolic blood pressure — are integrated into the model to reflect their combined influence on cardiovascular risk. The interaction between these factors was shown to significantly improve the predictive accuracy of the algorithm.

To leverage feature interactions in your own models, you don’t need to be working on healthcare algorithms. Interaction effects are everywhere — in marketing campaigns (where age and ad type might interact to influence customer behavior), finance (where income and credit score together shape lending decisions), and more.

When building models:

- Look for potential interactions in exploratory data analysis. Visualization tools like pairplots or interaction plots can help identify where relationships between features are more complex.

- Consider feature engineering, where you manually create interaction terms (e.g., by multiplying or combining features). Some algorithms, like decision trees or gradient boosting machines, can automatically account for interactions.

- Test your models with and without interactions to see the impact on accuracy and other performance metrics.