Section 3: Predictive Analytics for Forecasting and Optimization

17.3.1 The “Why”: The Shift from Reactive to Proactive

Thus far in this module, we have focused on using data to understand what has already happened. Descriptive analytics (the “what”) and diagnostic analytics (the “why”) are the bedrock of management. They allow you to report on your performance and explain the factors that drove it. This is akin to driving a car by looking exclusively in the rear-view mirror. You have a perfect, clear picture of the road you have already traveled, but you have no idea what lies beyond the next curve.

Predictive Analytics is the fundamental shift to looking through the windshield. It is the practice of using historical data, statistical algorithms, and machine learning techniques to identify the likelihood of future outcomes. This is the transition from a reactive manager—one who responds to problems as they arise—to a proactive, strategic leader who anticipates challenges and opportunities before they materialize. It is the single most powerful capability that separates good managers from truly great ones.

Imagine being able to answer these questions with a high degree of statistical confidence:

• What will our expenditures on IVIG be next quarter, and will it put us over budget?
• Which of our discharged heart failure patients are most likely to be readmitted within 30 days, and how can we intervene now?
• Based on surgical schedules and historical patient census, what is the mathematically optimal number of pharmacy technicians we need to have on duty next Tuesday at 2 PM to keep our turnaround times within target?

These are not questions of fantasy; they are the practical, high-value questions that predictive analytics allows you to answer. This section is designed to demystify these advanced techniques. We will demonstrate that predictive analytics is not a “black box” of incomprehensible algorithms, but rather a logical extension of the clinical forecasting you already do every day. You will learn the foundational methods for forecasting, the types of models used to predict outcomes, and the data required to power them. This is your introduction to becoming a truly forward-looking, data-driven pharmacy leader who shapes the future rather than simply reacting to the past.

17.3.2 The Analytics Maturity Model: A Roadmap to the Future

Organizations do not simply jump into predictive analytics overnight. They evolve their capabilities over time, moving through a well-defined continuum known as the Analytics Maturity Model. Understanding this model is crucial because it provides a roadmap for your department. It helps you identify where you are today, what the next logical step is, and what capabilities you need to develop to get there. The model progresses from looking at the past to shaping the future, with each level increasing in complexity and, more importantly, in the value it delivers to the organization.

17.3.3 Your Predictive Toolkit Part 1: Forecasting Fundamentals

Forecasting is the most accessible and immediately useful application of predictive analytics. It is the process of making predictions about the future based on past and present data, most commonly by analyzing trends. As a manager, you will constantly need to forecast: drug budgets, staffing needs, medication usage, and patient volumes. While perfect prediction is impossible, using formal forecasting methods can dramatically improve your accuracy and provide a data-driven rationale for your decisions. We will explore the most common and practical techniques you can begin using immediately.

The Foundation: Time Series Analysis

Most of the data you will forecast (drug usage, patient days, etc.) is time series data—a sequence of data points indexed in time order. To create a good forecast, you must first deconstruct the time series to understand its underlying components.

• Trend: The long-term direction of the data. Is your overall drug spend generally increasing, decreasing, or staying flat over several years?
• Seasonality: A predictable, repeating pattern that occurs over a known period (e.g., daily, weekly, yearly). The spike in RSV medication use every winter is a classic seasonal pattern.
• Cyclicality: A pattern that repeats over longer, less predictable timeframes. For example, hospital census might fluctuate with broader economic cycles over many years.
• Noise (or Random Variation): The unpredictable, irregular fluctuations in the data that are not explained by trend, seasonality, or cyclicality.

Method 1: Moving Averages (Smoothing Out the Noise)

The simplest forecasting method is the moving average. It works by taking the average of the last ‘n’ periods of data to forecast the next period. This is highly effective at smoothing out random noise to see the underlying trend.

Simple Moving Average (SMA): A 3-month SMA for September would be the average of June, July, and August. The formula is:

$$SMA_t = \frac{D_{t-1} + D_{t-2} + … + D_{t-n}}{n}$$

Where $D_t$ is the demand in period t, and n is the number of periods.

Weighted Moving Average (WMA): This is a more sophisticated version that gives more weight to more recent data, which is often more relevant. The formula is:

$$WMA_t = w_1 D_{t-1} + w_2 D_{t-2} + … + w_n D_{t-n}$$

Where the weights ($w_i$) must sum to 1.

Masterclass Example: Forecasting Albumin 25% Usage

You need to forecast the usage of Albumin 25% 100mL bottles for October. Here is your usage data for the past 6 months:

Month	Usage (Bottles)	3-Month SMA Forecast	3-Month WMA Forecast (Weights: 0.5, 0.3, 0.2)
April	210	–	–
May	230	–	–
June	225	–	–
July	245	(210+230+225)/3 = 221.7	(0.5*225)+(0.3*230)+(0.2*210) = 223.5
August	250	(230+225+245)/3 = 233.3	(0.5*245)+(0.3*225)+(0.2*230) = 236.0
September	265	(225+245+250)/3 = 240.0	(0.5*250)+(0.3*245)+(0.2*225) = 243.5
October Forecast	?	(245+250+265)/3 = 253.3	(0.5*265)+(0.3*250)+(0.2*245) = 256.5

Method 2: Regression Analysis (Finding the Relationship)

While moving averages are good for simple, trend-based forecasting, regression analysis is a much more powerful tool that allows you to forecast a variable based on its relationship with one or more other variables. It is the workhorse of predictive analytics.

The goal of regression is to find the mathematical equation that best describes the data. For a simple linear regression (one predictor variable), the equation is the familiar:

$$Y = mX + b$$

Where Y is the variable you want to predict (the dependent variable), X is the predictor (the independent variable), m is the slope of the line, and b is the y-intercept.

Masterclass Example: Forecasting Total Drug Spend

You know intuitively that as your hospital gets busier, your drug spend increases. Regression allows you to quantify this relationship and build a predictive model. Your dependent variable (Y) is “Monthly Drug Spend.” Your independent variable (X) is “Monthly Adjusted Patient Days (APD).” You gather the data for the last 12 months.

When you plot this data in a scatter plot and run a linear regression (a standard feature in Excel), you get a result like this:

What if you want to make your model even more powerful? You can use Multiple Linear Regression to add more independent variables. For example, you might find that surgical volume is also a major driver of drug costs. Your new model might be:

$$\text{Spend} = b_0 + m_1(\text{APD}) + m_2(\text{Surgical Cases})$$

This allows you to create even more nuanced and accurate forecasts by accounting for multiple business drivers simultaneously.

17.3.4 Your Predictive Toolkit Part 2: Classification and Optimization

While forecasting predicts a continuous value (like drug spend), another powerful branch of predictive analytics, called classification, predicts a categorical or binary outcome (e.g., Yes/No, High/Medium/Low). These models are used to classify individuals or events into groups, allowing you to target interventions with incredible precision. Beyond prediction, we will also introduce optimization, a prescriptive technique that doesn’t just predict the future but helps you find the best possible way to respond.

Classification Models: Answering “Will It Happen?”

A classification model analyzes a set of input variables (called “features”) to predict the probability that an event belongs to a certain class. This has transformative potential for proactive clinical pharmacy services and operational risk management.

Masterclass Example: Predicting 30-Day Hospital Readmissions

The Business Problem: Hospital readmissions are costly, harmful to patients, and heavily penalized by payers like CMS. Your hospital has tasked you with developing a program to reduce readmissions for patients with Congestive Heart Failure (CHF).

The Predictive Approach: Instead of providing the same level of intervention to all 200 CHF patients discharged each month, you want to use a classification model to identify the 20-30 patients who are at the highest risk of being readmitted. This allows you to focus your limited pharmacist resources where they will have the greatest impact.

Component	Description
The Goal	For every CHF patient being discharged, predict the probability that they will be readmitted to the hospital for any reason within 30 days.
The Model’s Output	A risk score (e.g., 0-100) or a simple classification (High Risk / Low Risk).
The “Features” (Input Variables)	You would work with a data analyst to extract historical data for thousands of past CHF patients. The model would learn the patterns from features like: Demographics: Age, gender. Clinical Data: LVEF (Left Ventricular Ejection Fraction), recent BNP lab value, creatinine clearance. Comorbidities: Presence of diabetes, COPD, renal failure (from ICD-10 codes). Utilization History: Number of ED visits in the last 6 months, number of prior hospital admissions. Medication Data: Number of home medications (polypharmacy), use of high-risk drugs like opioids or anticoagulants, prescription for a loop diuretic at discharge.
The “Label” (Historical Outcome)	For each patient in your historical dataset, you have a simple binary label: Was Readmitted in 30 Days (Yes/No)? The model learns the complex relationships between the features and this known outcome.
The Operational Intervention	The model runs automatically on every CHF patient scheduled for discharge. If a patient’s risk score exceeds a certain threshold (e.g., >75%), an automated alert is sent to the Transitions of Care pharmacist. This triggers a mandatory, high-intensity intervention: Bedside medication reconciliation and counseling. A follow-up phone call 48 hours post-discharge. Coordination with the patient’s outpatient pharmacy and cardiologist.

Optimization: Finding the “Best” Possible Answer

Optimization is a form of prescriptive analytics. It takes prediction a step further. While a predictive model might forecast your staffing needs, an optimization model will tell you the single best staffing schedule to meet your goals. It uses mathematical algorithms to find the optimal solution from a vast number of possibilities, given a specific objective and a set of constraints.

Masterclass Example: Optimizing the Technician Work Schedule

The Business Problem: You need to create the weekly schedule for your 20 pharmacy technicians. You have multiple, competing goals: you want to minimize your labor cost, but you also need to ensure that medication turnaround times are met, that technicians get the shifts they prefer, and that all hospital and union rules are followed.

The Predictive/Prescriptive Approach: An optimization model can solve this incredibly complex puzzle in minutes.

Component	Description
The Objective Function	This is the goal you want to achieve, stated mathematically. Most commonly: Minimize Total Labor Cost.
The Decision Variables	These are the things the model can change to achieve the objective. In this case, it’s a binary variable for each technician for every possible shift: Technician A works Shift 1 (Yes/No)? Technician A works Shift 2 (Yes/No)? …and so on for all techs and all shifts.
The Constraints	These are the rules or limits that the final solution must obey. This is where you build in your operational reality: Demand Constraints (from a predictive model): The number of technicians on duty during any given hour must be greater than or equal to the forecasted demand for that hour. Rule Constraints: Each technician must be scheduled for exactly 40 hours per week; each technician must have at least two consecutive days off; a technician cannot be scheduled for a day shift immediately following a night shift. Service Level Constraints: The resulting schedule must lead to a predicted average STAT turnaround time of less than 15 minutes (based on a simulation). Preference Constraints (soft constraints): The model can be designed to try and honor as many technician shift preferences as possible, adding a “satisfaction” component.
The Optimized Output	The model sifts through millions of possible schedule combinations and provides the single, mathematically optimal schedule that satisfies all constraints while achieving the lowest possible labor cost. It is a perfect, data-driven solution to a complex logistical problem.

17.3.5 Practical Implementation: The Data You Need and How to Get It

The sophisticated models described in this section are powerful, but they share a common, unyielding requirement: they are hungry for data. The quality, granularity, and accessibility of your data are the ultimate limiting factors in your ability to leverage predictive analytics. The “Garbage In, Garbage Out” principle we’ve discussed before applies with even greater force here; a predictive model trained on flawed, incomplete, or biased data will produce flawed, incomplete, or biased predictions. A key part of your role as a data-driven leader is to understand the data ingredients required for your desired analytical outcomes and to champion the initiatives needed to capture and organize that data effectively.

Masterclass Table: Predictive Model Data Ingredients

This table outlines the data requirements for the predictive models we have discussed. It is designed to help you understand the types of data you will need to start collecting and organizing to enable these advanced capabilities.

Predictive Goal	Key Data Sources Required	Specific Data Elements (“Features”) Needed	Common Hurdles & How to Overcome Them
Forecast Monthly Spend for a High-Cost Drug (e.g., IVIG)	PIS / Wholesaler Purchasing History EHR Medication Administration Record (MAR) ADT (Admit-Discharge-Transfer) System	Historical purchase dates and quantities Historical administration dates and doses Historical patient census / patient days (Advanced) Number of patients with specific ICD-10 codes (e.g., primary immunodeficiency)	Hurdle: Distinguishing between usage (administrations) and purchases. A large purchase in one month can skew the data. Solution: Base your forecast on the MAR administration data, as it is a truer reflection of patient demand. Use purchasing data to validate your model.
Predict Patient Risk of 30-Day Readmission	EHR Clinical Data Repository ADT System Pharmacy Information System (PIS)	Patient demographics (age) Diagnosis codes (ICD-10) for comorbidities Lab values (e.g., BNP, SCr) Prior utilization (ED visits, admissions in last 6 months) Number of discharge medications	Hurdle: Data lives in multiple, separate systems. Solution: This project requires a strong partnership with your IT/data analytics department. They will need to create a unified, patient-level dataset (often called a data warehouse or data mart) that brings together clinical, demographic, and pharmacy data.
Optimize Technician Staffing Schedule	Time-stamped data from PIS and EHR Employee timeclock/scheduling system ADT System	Hourly counts of new orders, verified orders, dispensed doses (from ADCs, carousels), and IVs compounded. Historical technician clock-in/clock-out data. Historical patient census by nursing unit. Technician skill mix (e.g., certified IV tech vs. general tech).	Hurdle: Capturing granular, time-stamped workload data can be difficult. Solution: Leverage the data that is already being captured. Your PIS and ADC systems log the exact time of every transaction. Work with your vendors or IT to get access to these raw transaction logs, which are a goldmine of workload data.