Machine learning excels at prediction.
Given sufficient data, a model can identify complex relationships between patient characteristics and clinical outcomes.
However, predictive relationships should not automatically be interpreted causally.
Consider a model that identifies treatment exposure as an important predictor of poor outcomes. This does not necessarily mean the treatment caused those outcomes. Patients receiving the treatment may have been systematically different from patients who did not receive it.
This is a classic example of confounding.
The distinction becomes important when healthcare AI moves from risk estimation toward intervention.
A predictive model can estimate an individual's probability of an outcome. A causal framework is concerned with how that outcome might change under a defined intervention or counterfactual scenario.
These are fundamentally different questions.
Model development should therefore begin with a clear definition of the decision problem.
If the objective is prediction, appropriate predictive evaluation is required.
If the objective is estimating treatment effects or choosing between interventions, causal assumptions and methods become central.
Causal inference may involve randomized experiments, target trial emulation, propensity-based methods, instrumental variables, longitudinal approaches, or other frameworks depending on the problem and available data.
The important point is conceptual.
A feature with high predictive importance is not automatically a useful intervention target.
A strong association is not automatically evidence that changing the associated variable will change the outcome.
As healthcare AI becomes more involved in treatment selection and personalized interventions, maintaining this distinction will be essential.
Good healthcare AI should know whether it is predicting what will happen or estimating what could happen if we act.
Top comments (0)