Modeling in health is the use of mathematical equations, computer simulations, and statistical tools to represent how diseases spread, how treatments work, and how health policies will play out before they’re implemented in the real world. It’s essentially a way of running experiments that would be impossible, unethical, or too slow to conduct on actual populations. Health models range from simple equations tracking an infection through a community to complex artificial intelligence systems predicting which patients are most likely to end up in the hospital.
How Health Models Work at a Basic Level
Every health model starts with the same idea: take what we know about a disease, a treatment, or a population and translate it into a set of rules that a computer can process. Those rules might describe how quickly a virus passes between people, how a drug breaks down in the body, or how likely someone with certain risk factors is to develop heart disease. The model then runs those rules forward in time, generating predictions about what will happen under different conditions.
Models fall into two broad categories. Deterministic models follow fixed rules and produce a single predicted outcome every time you run them. If you plug in the same starting numbers, you get the same result. Stochastic models introduce randomness, reflecting the reality that biology is unpredictable. Running a stochastic model multiple times generates a distribution of possible outcomes rather than one answer, which tends to produce more realistic predictions. Disease transmission, for example, is inherently random, so adding environmental noise to a model captures the kind of real-world variability that a purely deterministic approach misses.
Tracking Infectious Disease Outbreaks
The most widely known type of health modeling is the compartmental model used in epidemiology. The simplest version, called the SIR model, divides an entire population into three groups: Susceptible (people who can catch the disease), Infected (people who currently have it and can spread it), and Removed (people who have recovered or died and are no longer part of the transmission chain). The model uses two key numbers to drive everything: the transmission rate, which captures how many people a single infected person exposes per day, and the recovery rate, which captures how quickly people get better.
From those two numbers, the model calculates the basic reproduction number, often written as R₀. This is the average number of new infections caused by one sick person in a fully susceptible population. When R₀ is above 1, an outbreak grows. When it drops below 1, the outbreak fades. More complex versions add stages. The SEIR model, for instance, inserts an “Exposed” group between Susceptible and Infected to account for the delay between catching a pathogen and becoming contagious. Other variations add categories for people who lose immunity over time and become susceptible again.
These models were central to the COVID-19 response. In Jordan, scenario-based modeling compared four physical distancing strategies during 2020 and 2021, ranging from no restrictions to permanent closures of nonessential services. The models showed that permanent distancing would reduce cases and deaths most effectively, but only marginally more than intermittent distancing (closing nonessential services one or two days per week). That finding directly shaped the government’s policy: rather than imposing a full lockdown, Jordan adopted intermittent restrictions with a six-hour overnight curfew on other days. The actual course of the pandemic in Jordan confirmed the model’s forecasts.
Predicting Individual Patient Risk
Not all health modeling operates at the population level. Risk stratification tools use statistical or machine learning techniques to generate individual risk scores, predicting outcomes like hospital admission, emergency department visits, healthcare costs, or death. Hospital admission is the most commonly predicted outcome in these models, followed by healthcare costs and emergency department attendance.
These tools pull from a wide range of data. Electronic health records provide demographics, medications, vital signs, lab results, and visit histories. Insurance claims data add information about procedures, office visits, and hospitalizations. Increasingly, patient-generated health data from wearable devices, smartphone apps, and remote monitoring tools feed into these models as well. A 2017 survey found that 68 percent of patient registries in the United States already extracted some data from electronic health records, and that number has only grown.
The best machine learning models in this space have achieved strong accuracy scores (C-statistics as high as 0.90, where 1.0 would mean perfect prediction and 0.5 would mean random guessing). But accuracy alone doesn’t guarantee usefulness. One well-known risk tool, the Predictive Risk Stratification Model (PRISM), was linked to an intervention that actually increased hospital admissions by 44 percent, along with more emergency visits and greater primary care workload. A different tool, the Nairn Case Finder, reduced hospital admissions by 42.5 percent among high-risk patients in a smaller study. The difference wasn’t just in the model itself but in how the predictions were used to guide care.
Evaluating Whether Treatments Are Worth the Cost
Health economic modeling helps governments and insurers decide which treatments to fund. The most common approach uses a structure called a Markov model, which maps out all the health states a patient could move through over time (healthy, mild disease, severe disease, death) and assigns each state a cost and a quality-of-life value. Quality-of-life is measured in QALYs, or quality-adjusted life years, where one QALY equals one year lived in perfect health.
By simulating thousands of patients moving through these health states under two different treatments, the model produces total costs and total QALYs for each option. The key output is the incremental cost-effectiveness ratio (ICER), calculated by dividing the difference in cost between two treatments by the difference in QALYs gained. If a new drug costs $29,000 more than the standard treatment but delivers an extra 165 QALYs across a population, the ICER works out to about $177 per QALY gained. Policymakers then compare that number to a threshold (often $50,000 to $150,000 per QALY in the U.S.) to decide whether the treatment represents good value.
Modeling How Drugs Move Through the Body
In drug development, pharmacokinetic-pharmacodynamic modeling tracks what happens after someone takes a medication. The pharmacokinetic side describes how the body absorbs, distributes, and eliminates the drug. The pharmacodynamic side describes what the drug does once it reaches its target. Together, these models connect a drug’s concentration in the blood to its actual effects.
This matters because the same dose of a drug can produce very different blood concentrations in different people, depending on their weight, organ function, genetics, and other medications. By modeling the relationship between concentration and effect rather than simply dose and effect, researchers can identify why a drug works well in some patients and poorly in others. These models also help translate findings across species, making it possible to interpret results from animal studies in terms of what will likely happen in humans.
Simulating Communities With Agent-Based Models
Agent-based modeling takes a fundamentally different approach from the equation-based methods described above. Instead of treating a population as a set of averages, it creates a virtual world populated by individual “agents,” each with their own characteristics, behaviors, and social connections. These agents interact with each other and their environment according to programmed rules, and population-level patterns emerge from the bottom up.
Each agent can differ in age, income, health status, social network, and behavior. They make autonomous decisions based on their circumstances. A simulation of obesity, for example, might give each agent a neighborhood with certain food options, a social network that influences eating habits, and a set of biological parameters governing weight gain. Running the simulation reveals how neighborhood-level changes (like adding a grocery store) ripple through social networks to affect population health in ways that wouldn’t be obvious from studying individuals in isolation. These models capture feedback loops, where past experiences change future behavior, and nonlinear effects, where small changes can amplify into large outcomes over time.
Bias and Transparency Challenges
Health models are only as good as the data behind them, and that data often reflects existing inequalities. Chest X-ray classifiers trained on large public datasets have been shown to underdiagnose conditions in underserved racial groups. An algorithm designed to detect acute kidney injury underperformed for women because the training data skewed heavily male. Without careful examination of how biases get encoded into these tools, there is a real risk of scaling up health disparities rather than reducing them.
The STANDING Together initiative, published in The Lancet Digital Health, calls for transparency in health datasets: honest reporting of limitations, proactive evaluation of how models perform across different population groups, and clear documentation of what a dataset can and cannot represent. For anyone whose care is influenced by a predictive model, these aren’t abstract concerns. A model trained primarily on data from one demographic group may generate misleading risk scores for everyone else.