Risk Stratification and Prediction

The first thing an analytics function is asked to do is find the patients who need attention. This is risk stratification, sorting a population by expected need, and the predictive models behind it. Done well, it points care management where it can help; done naively, it chases patients who would have been fine anyway.

Stratification, from data

The population-health pyramid from earlier courses is, in data terms, a ranking problem. The analytics function computes a risk measure for each patient and sorts them into tiers: the high-need apex, the rising-risk middle, and the stable base. The risk measure can be as simple as prior cost or as sophisticated as a predictive model combining diagnoses, utilization, labs, medications, and social factors.

From prediction to impactability

Here the evaluation course’s warning becomes an engineering principle.

Worth remembering: a model that predicts who will be high-cost is not the same as a model that finds who you can help. Last year’s most expensive patients partly regress to the mean on their own, so targeting them alone produces apparent success that is really arithmetic. The higher-value analytic target is impactability: which patients are on a worsening trajectory that intervention can actually bend. Building or buying models that estimate impactability, not just cost, is what separates analytics that generate real savings from analytics that generate flattering reports.

What goes into a risk model

Modern risk and predictive models draw on the full data landscape:

• Diagnoses (the HCC-style condition categories from the risk-adjustment course).
• Utilization patterns (prior admissions, ED visits).
• Clinical values (labs, vitals) where available.
• Medications (adherence, complexity).
• Social factors (area-level or screened).

The richer and better-integrated the data, the better the model, which is why the pipeline and quality work of Module 3 pays off directly here.

The honest caveats

Predictive models carry the cautions the whole curriculum has raised. They predict weakly at the individual level and work best in aggregate. They inherit the biases of their training data, including the equity concerns from the intro course, so a model built on historically unequal care can perpetuate it. And a model is only as good as the action it triggers: a perfect risk list that no one works changes nothing. Analytics teams that treat models as decision support for human care managers, not as autonomous truth, use them best.

Key takeaways

• Risk stratification ranks a population by expected need, from simple prior-cost measures to sophisticated predictive models.
• Predicting impactability, who intervention can help, beats predicting cost alone, which invites regression to the mean.
• Models draw on the full integrated data landscape, predict weakly for individuals, can inherit bias, and only matter if they drive action.