AUSTIN, TEXAS - Artificial intelligence (AI) applications in healthcare are improving clinical decision making and therapies and enhancing clinical and operational efficiencies. They offer the potential to radically transform medical practice and improve global health. While this is great news, the misapplication of AI models can easily lead to undesirable results: To avoid this, the formulation and application of AI models require different philosophical and mathematical assumptions.

This article focuses on revealing different model assumptions - specifically the difference between predictive and prescriptive models - which is a greatly misunderstood topic for AI in healthcare.

Correlation versus causation

In every Statistics 101 class, students learn that ‘correlation is not causation.’ With good reason, one of this author’s favorite quotes is “if you torture the data long enough, they will admit to anything.” In other words, if one compares enough pairs of variables, one will inevitably find some with high degrees of correlation. However, this is happenstance correlation, and it is the worst type of correlation when it comes to AI modeling. Why? The answer is that these variables simply have nothing in common from a logical standpoint. Some examples from author, military analyst, and Spurious Correlations website creator Tyler Vigen:

Correlation between US spending on science vs. suicides: 99.79 percent (yearly, 1999 to 2009)

Correlation between the number of people who drowned after falling out of a fishing boat vs. the marriage rate in Kentucky: 95.24 percent

One can call these ‘useless correlations’ for AI: such correlations do not provide any useful insights, nor is there any causative relationship.

Then there are other relationships where there is a rati