If a computer model predicts how blood flows through a heart, which assumption should you question first? A model may contain many inputs, from material stiffness to geometry and boundary conditions. Sensitivity analysis answers by changing assumptions in a controlled way and tracking how much the predicted result moves. It does not tell us which answer is true by itself; it shows where the model’s conclusions depend most strongly on what was assumed.
A model is a chain of choices
Computational modeling uses mathematics, physics, and computer science to simulate complex systems. A model represents a system through variables, equations, and assumptions, then produces outputs under specified conditions [1]. For a heart-flow model, these choices might include how the valves are represented, the properties assigned to tissue, and the conditions used where blood enters and leaves the simulated region.
Sensitivity analysis begins by identifying inputs that could plausibly vary. One can imagine turning a single dial while holding the others fixed: if the predicted flow changes sharply, that input is influential for that output. Then other dials can be varied. This isolates cause and effect within the model, much like changing one condition in a physical experiment, though the result remains a statement about the model’s behavior [1].
Separate influence from uncertainty
An influential assumption is one whose variation changes the output substantially. An uncertain input is one whose value is not known precisely. These are related but distinct. A poorly known parameter may have little effect on the prediction; another parameter may be well measured yet strongly control the result. Sensitivity analysis helps identify which uncertainties deserve attention first.
Measurement uncertainty is a formal part of science, not a synonym for error or ignorance. NIST describes methods for evaluating uncertainty components and combining them to express uncertainty in measurement results [2]. If a model uses measured quantities, those uncertainty ranges can inform which input variations to explore. The key is to distinguish variation in the real system from uncertainty in how a quantity was measured.
A small thought experiment
Suppose a model predicts the flow through a valve. First vary the valve opening slightly while keeping fluid properties and pressure conditions fixed; then restore it and vary the assumed tissue stiffness. If one change moves predicted flow much more than the other, the first input has greater influence in that tested range. The result depends on the output chosen: the same input might strongly affect peak flow but barely alter total volume.
This simple one-at-a-time picture is useful, but inputs can interact. A model can behave differently when two assumptions change together than when either changes alone. Biomedical models may combine mechanistic descriptions based on physical or biochemical principles with data-driven patterns, and they can span scales from molecules to tissues and organisms [1]. Careful analysis explores the combinations relevant to the question rather than treating every input as independent.
What the analysis can and cannot establish
If a model’s prediction barely changes across plausible input ranges, that output is relatively robust to those assumptions. If it changes substantially, the result is fragile and further measurement or model refinement may be valuable. Sensitivity analysis can therefore guide experiments and help prioritize data collection, but it cannot verify that the model’s structure correctly represents reality. A stable answer can still be wrong if an important mechanism was omitted.
NIST’s framework for uncertainty provides ways to evaluate and combine contributions to uncertainty, while computational models allow simulated conditions to be compared systematically [2] [1]. Together, these ideas help make a prediction more transparent: readers can see which assumptions drive it, how measurement limits enter, and where evidence is still needed.
Sources & further reading
Prepared for the independent PSI Center reading library. About edition dates and source checking.
This explainer describes scientific principles. It does not assess individual treatment choices.
