AUSTIN, TEXAS - Part 1 of this series covered the classification of artificial intelligence (AI) and machine learning (ML) models in medicine - the analytics maturity ladder. The most powerful AI models in medicine answer the ‘what’ and ‘why’ questions. Examples of ‘what’ questions are:

What disease is present?

What happened?

What is the likelihood that this event will happen?

In these circumstances, a clinician or scientist wants to know how to understand the underlying mechanisms that produce an outcome, which is certainly useful. However, there is a more important set of questions that are more difficult to answer - the ‘why’ questions:

Why did this happen? 

How can one make it happen?

What is the best therapy to follow for this patient?

These are very different questions and require a different level of mathematical rigor. Burk and Miner 2020 explain that all predictive modeling is based on statistical modeling. Statistical models are simply trying to understand the characteristics of a population of interest, which often contains unforeseen future outcomes. To model or understand these unseen observations, one collects samples and generates mathematical characterizations called statistics. Any function of a sample is a statistic.

This article offers a very high-level overview of some bridges from statistics to causal models. These models are useful in unders