In a past post, I mentioned a probabilistic framework for optimal medical decision making. This is a more in-depth discussion of the framework, along with some difficulties that arise when using it.

To summarize the framework in words:

If we can assign a utility, or a degree of “goodness,” to all possible outcomes, e.g.,

1. Disease cured, no side effect

2. Disease not cured, no side effect

3. Disease cured, side effect

4. Disease not cured, side effect

And we can also determine the probability of each outcome under treatment and non-treatment, e.g.,

1.a. Probability Disease cured, no side effect under treatment

1.b. Probability Disease cured, no side effect under non-treatment

2.a. Probability Disease not, no side effect under treatment

2.b. Probability Disease not, no side effect under non-treatment etc…

Then we can choose the decision (treat, or don’t treat) that gives the most utility on average.

It sounds reasonable. But, then, why isn’t this framework used every time we go to the Doctor’s office?

1. Assigning utilities is difficult.

How do we choose the scale, for example?

How do we ensure that our numbers reflect the way the patient truly feels?

How to discuss these outcomes in a sensitive way?

2. Determining the probabilities of the outcomes is difficult.

These can be “imputed” with clinical intuition.

These can also be estimated from trial data.

These are sometimes dependent on patient covariates, which are not always collected.

To read more, see e.g., Pauker’s review on medical decision analysis.

Pauker, Stephen G., and Jerome P. Kassirer. “Decision analysis.” New England Journal of Medicine 316.5 (1987): 250-258.