Individual Patient Simulation

Individual Patient Simulation: Concept Architecture

Orientation and learning roadmap

The page follows simulated patients from baseline characteristics through event histories to costs and health outcomes. It compares those histories with a cohort average, checks a small random-draw example, and separates patient-level randomness from uncertainty in model parameters. The last sections explain model choice, precision, validation, and reporting.

Specify patients, histories, and decisions

An individual-level model assigns each simulated person characteristics such as age, disease severity, or previous events, then updates that person's trajectory under a specified strategy. Risks, treatment eligibility, costs, and utilities can depend on the current characteristics and accumulated history. Repeating this process over a target population gives a distribution of individual outcomes and estimates of population means.

The model can use discrete cycles, exact or sampled event times, or other suitable mechanisms. Patient-level simulation describes the level at which histories are represented; it is not synonymous with a particular time engine such as discrete-event simulation. Interactions among people or shared resource constraints require explicit additional structure rather than arising automatically from individual records.

Run events and accumulate outcomes

At each relevant step, the model calculates event risks or event times from the patient's state and the selected parameters. It draws stochastic events when appropriate, updates the person's history, and adds costs or health outcomes when they occur or accrue. Stopping rules, competing events, and discounting must be consistent across strategies.

For $N$ simulated people, an estimated mean outcome under one parameter set is

$$\overline C=\frac{1}{N}\sum_{i=1}^{N}C_i, \qquad \overline Q=\frac{1}{N}\sum_{i=1}^{N}Q_i,$$

where $C_i$ is a person's cumulative cost and $Q_i$ the person's cumulative QALYs over the stated horizon. The arithmetic averages completed trajectories; the rules generating those trajectories are the substantive clinical model. When people are drawn independently from the target population, simulation precision generally improves with more draws, but biased or misspecified rules do not become valid merely by increasing $N$.

Check a three-patient illustration

Suppose the first-year event probabilities for three illustrative people are 0.10, 0.30, and 0.50, and each receives an independent uniform random draw. An event occurs when the draw is below that person's probability. Everyone has utility 0.8 during year one; an event at the year-one boundary lowers year-two utility to 0.7, while no event leaves it at 0.8.

The example assigns £100 in baseline cost per person and another £500 to an event, with no discounting or mortality. The realised mean is $(600+100+600)/3\approx£433.33$ and the realised mean QALYs are $(1.5+1.6+1.5)/3\approx1.533$. Given these three fixed probabilities and model rules, the probability-weighted expected values are instead £250 and 1.57 QALYs per person; this three-draw result differs because of simulation noise.

The calculated expectation follows from the mean event probability $(0.10+0.30+0.50)/3=0.30$. It gives $100+500(0.30)=£250$ and $1.6-0.1(0.30)=1.57$ QALYs. Three people are far too few to estimate a stable decision result, but the discrepancy clearly separates a simulated sample average from the model's expectation.

Distinguish sources of variability and uncertainty

Different patients have different characteristics and outcomes even if all input parameters are fixed. Stochastic trajectory draws add first-order Monte Carlo variation to a finite simulation estimate; repeated runs or more simulated people can measure and reduce its numerical effect. Uncertainty about treatment effects, transition parameters, costs, or utilities is a separate parameter-uncertainty layer explored through probabilistic sensitivity analysis (PSA).

In a nested PSA, each outer draw samples a coherent set of uncertain parameters and the inner simulation estimates strategy outcomes conditional on those parameters. Inner simulation error must be small enough that it does not materially distort the distribution of economic results across outer draws. Structural uncertainty, such as an omitted disease mechanism, remains a further issue that more patient draws cannot resolve.

Compare strategies efficiently and fairly

Strategy comparisons should represent the same target population and preserve relevant correlations among a person's characteristics. Using common random numbers across alternatives can reduce Monte Carlo noise in their difference when corresponding event mechanisms are coupled sensibly. The precision of the incremental result depends on the variability of paired differences, not only the precision of each arm's mean.

An individual-level approach can be justified when risk depends on continuous characteristics, recurrent events, past treatment, or complex eligibility rules that would make a cohort state space unwieldy. A cohort model may be simpler and equally suitable if aggregated states capture what matters. Extra detail has a cost in data needs, computing time, debugging, and explaining results.

Validate histories and reported outputs

Verification checks whether simulated events obey the intended rules and whether costs and QALYs accumulate once at the correct times. Validation compares predictions with appropriate observed outcomes, independent data, and clinical knowledge for the intended population. Neither good face validity nor a visually plausible individual trajectory is enough to validate population cost-effectiveness.

• Inspect fixed-seed patient traces for impossible state transitions, duplicate costs, and incorrect stopping times.
• Compare simulated event frequencies with their input probabilities in a simple test case.
• Reconcile simulated population means with analytic expectations where a reduced model permits it.
• Report simulated patient counts, replications, seeds, Monte Carlo error, and PSA design.
• Check baseline characteristics, event rates, and survival against relevant data and assess target-population transferability.
• Test whether increasing simulation size changes incremental costs, QALYs, or the preferred strategy materially.

Sources and technical basis

The NICE Decision Support Unit describes design, implementation, and reporting of cost-effectiveness modelling using patient-level simulation. ISPOR-SMDM guidance distinguishes individual state-transition microsimulation and uncertainty types; methodological work addresses Monte Carlo error when PSA is combined with patient simulation. The three-patient example is original and is deliberately too small for inference.

• NICE Decision Support Unit, Cost-Effectiveness Modelling Using Patient-Level Simulation, TSD 15.
• ISPOR-SMDM, State-Transition Modeling, good research practices.
• ISPOR-SMDM, Model Parameter Estimation and Uncertainty Analysis.
• O'Hagan and colleagues, Monte Carlo probabilistic sensitivity analysis for patient-level simulation.