Problem Formulation Before Model Selection in Healthcare AI

A common machine learning workflow begins with a dataset and a prediction target.

Define the label.

Select features.

Train models.

Tune hyperparameters.

Evaluate performance.

But healthcare AI often requires an earlier step:

Define the problem correctly.

Not every clinical or operational problem should be formulated as prediction.

Prediction estimates the probability of an outcome.

It does not automatically identify the best intervention.

This distinction is important.

Suppose a model predicts that a patient has a high probability of readmission.

The prediction may be accurate.

But if the clinical team does not know which intervention will reduce that risk, the predictive model may have limited decision value.

Similarly, predicting disease progression is different from estimating the causal effect of an intervention on that progression.

Different questions require different methods.

A practical healthcare AI problem formulation should clarify the target decision, decision-maker, information available at decision time, action that follows, consequences of errors, and appropriate evaluation criterion.

The technical solution can then be selected accordingly.

A prediction model may be appropriate for risk stratification.

Information retrieval may be more appropriate when the problem is locating relevant evidence across a longitudinal record.

Summarization may help when the challenge is information overload.

Optimization may be appropriate for scheduling or resource allocation.

Causal inference may be required when the question concerns intervention effects.

Automation may be sufficient for repetitive workflows.

Even rules-based systems can remain appropriate for clearly defined deterministic processes.

This matters for model evaluation.

If the actual objective is to improve a decision, AUROC alone may not be enough.

We may need to evaluate decision quality, clinical outcomes, workflow impact, calibration, resource utilization, or other measures aligned with the intended use.

The central principle is simple:

Choose the computational method based on the healthcare problem, not the other way around.

A technically sophisticated solution to the wrong problem is still the wrong solution.