Section 12.3: Predictive Analytics for Risk Stratification

12.3.1 The “Why”: The Clinical & Financial Failure of Reactive Care

In the previous section, we built a powerful dashboard. This tool gives you an exceptional, at-a-glance view of your pharmacy’s performance. It tells you, with data, what your “gut feeling” used to: when you are busy, when quality is slipping, and where your bottlenecks are. This is descriptive analytics—it describes the past and the present. It is a massive leap forward, but it is still, fundamentally, a reactive tool.

You look at your dashboard and see “Patient John Smith, PDC for Eliquis: 45%.” You have successfully identified a non-adherent patient. But what does this mean? It means the patient has already been non-adherent for weeks or months. The harm—be it disease progression or an increased risk of stroke—has already begun. The “late to refill” report, the core of most pharmacy workflows, is a list of failures. You are a highly-skilled clinical detective arriving at the scene after the crime has been committed.

This reactive model is a clinical, financial, and moral failure:

• It’s Clinically Ineffective: You are intervening after the patient has failed therapy. You are trying to reverse months of poor adherence, which is infinitely harder than preventing it from the start.
• It’s Financially Wasteful: This model is the definition of “a day late and a dollar short.” The PBM has already tagged this patient as non-adherent and is already planning to penalize you with DIR fees. The health plan is already exposed to the massive cost of a potential ER visit or hospitalization that you failed to prevent.
• It Wastes Your Resources: Your “late to refill” list is long and full of “noise.” You spend hours calling 100 patients. 80 of them are “low-risk” (e.g., they just forgot and will pick it up tomorrow, or they went on vacation). 20 are “high-risk” (e.g., they stopped the med due to a severe side effect, or they can’t afford the copay). You gave equal effort to all 100, meaning your most valuable, high-touch interventions were diluted.

Predictive Analytics is the solution to this problem. It is the practice of using your rich Real-World Data (RWD) to forecast the future. Instead of looking at a dashboard that tells you “who is non-adherent,” you will build and use a tool that tells you “who is most likely to become non-adherent in the next 30 days.”

This shift moves you from a reactive detective to a proactive population health manager. You can now focus your limited, high-value pharmacist time only on the 20 high-risk patients, and intervene before they ever miss a dose. You stop calling the 80 low-risk patients, saving valuable time. This is Risk Stratification: the process of “bucketing” your patient population by their predicted risk, so you can apply the right intervention, to the right patient, at the right time. This is the single most valuable, and most advanced, application of data in modern pharmacy practice.

12.3.2 Pharmacist Analogy: Your “Clinical Gut Feeling” is Already a Predictive Model

12.3.3 How a Predictive Model is Built: A 5-Step Guide for Pharmacists

You do not need to be a data scientist or a statistician to use a predictive model. However, as a CASP, you must be an intelligent “consumer” of these models. You need to understand how they are built so you can assess their quality, identify their biases, and, most importantly, trust their outputs. The process can be broken down into five core steps.

12.3.4 Masterclass Table: Common “Target Variables” to Predict in Pharmacy

The first step is deciding what to predict. Your choice of “target” will define your entire program. A model built to predict non-adherence is useless for predicting ADEs. Below are the most common and high-value targets for a CASP.

12.3.5 Masterclass Table: The Feature Engineering “Cookbook”

This is your “cookbook” of ingredients. A predictive model is only as smart as the features you feed it. As a pharmacist, your greatest value is recommending these features. A data scientist may not know that a “vacation override” is a sign of chaos, or that a `Depression` diagnosis is one of the single highest predictors of non-adherence. You do. Your clinical insight is what makes the model smart.

12.3.6 A Pharmacist-Friendly Guide to Model Types (Interpreting the “Black Box”)

The algorithm itself can seem like a “black box.” While you don’t need to code the algorithm, you need to understand the logic of how it “thinks.” This helps you trust the output and explain it to stakeholders. The two most common and useful model types in pharmacy are Logistic Regression and Decision Trees.

Model 1: Logistic Regression (The Workhorse)

This is the most common, most trusted, and most interpretable model. It’s used when your “target” is binary (Yes/No). It is simply a formula that calculates the probability (from 0% to 100%) that your patient will have the “Yes” outcome. Its output is a list of “Odds Ratios” for each feature, which are incredibly easy for a clinician to understand.

Model 2: Decision Trees / Random Forests (The Intuitive Flowchart)

This model works exactly how your brain does. It’s a series of “If-Then” questions that splits the population into smaller and smaller buckets. A “Random Forest” is simply a “forest” of hundreds of different “trees” that all vote on the final prediction (this makes it much more accurate and stable).

This is the easiest model to visualize and explain to your staff.

Visual Tutorial: A Simple Decision Tree for Predicting Non-Adherence

Imagine this flowchart is your algorithm. You start at the top with a new patient and follow the “Yes/No” path down to their final risk bucket.

12.3.7 The Grand Finale: Building the Risk-Stratified Workflow

You have a validated model. Every night, it runs and generates a “Risk Score” from 0 to 100 for all 10,000 of your patients. This is the moment of truth. How do you turn this list of 10,000 numbers into an actual, efficient workflow? You do it by creating risk-stratified “buckets” and defining a different intervention playbook for each.

This is the single most important concept: Stop treating all patients the same. Your “white glove” pharmacist-led interventions are expensive and scarce. They must be saved only for the patients who need them the most. Conversely, your “low-touch” automated reminders are cheap and scalable, but they are ineffective for high-risk patients. You must match the intensity of the intervention to the intensity of the risk.

Masterclass Table: The Risk-Stratified Intervention Playbook

12.3.8 The Ethical Imperative: Bias, Fairness, and How to Trust Your Model

A predictive model is a powerful tool. But like any tool, it can be dangerous. A model is an amplifier. If it is built on a foundation of biased data, it will not just replicate that bias; it will amplify it and give it the false authority of “objective data.” As a CASP, you are the last line of defense. You are the clinical and ethical guardian who must question the “black box.”

Correlation is NOT Causation:
Your model might find that patients who use your “auto-refill” service are more adherent. This is a correlation.
The Wrong Conclusion (Causation): “Auto-refill causes adherence! We should just enroll everyone!”
The Right Conclusion (Correlation): “Patients who are already organized, motivated, and adherent are the ones who sign up for auto-refill. They are two related effects of a third, unmeasured variable (the patient’s ‘motivation’).” This is called “confounding,” and it’s why you must be a skeptical, clinical expert.

How Do We Know a Model is “Good”? (A 2-Minute Guide to Validation)

When your data team “validates” the model (Step 4), they will give you a report. You don’t need to understand all the math, but you need to understand the two main “scores,” which are identical to how you think about diagnostic tests.