Most medical AI is trained to look backward. It learns from patients who have already been observed and predicts a label: deterioration, readmission, treatment response, or diagnostic probability.

But many of medicine’s hardest questions are not prediction questions.

What happens if the dose changes? Which tissue should be ablated? How would a device perform in an anatomy that was rare in the original trial? What if a hospital changes its staffing or triage policy?

These are intervention questions. Historical data alone contain only the actions that were actually taken. Simulation adds something different: a controlled world in which alternative actions can be tried before they reach a patient.

The next important layer of medical AI may not be another predictor. It may be an environment in which predictions can be turned into experiments.

The Earlier Form of Simulation

Traditional simulation begins with a stated mechanism. In a classroom flu model, if 60 classmates each independently faced a 1% infection probability, expected first-day infections were 60 × 0.01 = 0.60. If repeated runs did not approach that value, the implementation or the stated assumptions needed review.

This is still the right starting point: check simple expectations, edge cases, and conservation rules before trusting larger outputs. Such models are transparent but usually generic; medical AI can help personalize or update them.

Simulation Is Already Part of Medical Evidence

This direction is not speculative. Pharmacokinetic and physiologically based models already connect dose, exposure, efficacy, and safety; the FDA’s final ICH M15 guidance, issued in June 2026, sets out a framework for model-informed drug-development evidence. For devices, a risk-informed credibility framework addresses mechanistic simulations.

VICTRE created virtual breast anatomies, inserted lesions, simulated image acquisition, and compared two imaging modalities in an in silico trial. Across drug, device, workflow, and patient-specific models, the goal is the same: test a defined question in a controlled representation of the system.

AI and Simulation Solve Opposite Problems

Traditional simulation is strongest where mechanisms are known; machine learning is strongest where data are abundant. A hybrid model combines a mechanistic prediction with an AI-learned residual: patient data personalize parameters, AI fills limited gaps or accelerates solvers, and the mechanism constrains states and interventions. A published pharmacokinetic example combines deep learning for molecular properties with a physiologically based model for concentration-time profiles.

From Population Models to Patient-Specific Experiments

flowchart LR
    A["Population simulation"] --> B["Virtual cohort"]
    B --> C["Patient-specific model"]
    C --> D["Digital twin"]
    D --> E["Closed-loop learning system"]

A population simulation varies assumptions; a virtual cohort varies simulated patients; a patient-specific model is calibrated to one person; and a digital twin also updates with observations. A closed-loop system recalibrates from resulting actions.

In a 2026 10-person feasibility study, heart models identified ventricular-tachycardia ablation targets before procedures. Ventricular tachycardia was noninducible after ablation in all participants; at mean 13-month follow-up, eight were recurrence-free without antiarrhythmic medication. It is a step toward treatment planning, not evidence for routine use.

A Plausible Synthetic Patient Is Not Yet a Simulator

Generative AI can produce realistic-looking records, trajectories, and patient profiles for education, testing, or data augmentation. But realism is not intervention validity: a language model may imitate laboratory results after a medication change without representing the processes that connect drug and outcome.

For a simulator to support a medical decision, it needs a defined context of use, a defensible intervention mechanism, calibration to the represented patient or population, uncertainty that grows outside its evidence base, and validation against decision-relevant outcomes.

This is the central boundary for AI-era simulation:

A model becomes medically useful not when it can generate a believable future, but when it can distinguish which futures its evidence allows it to compare.

The First Equation Still Matters

The flu model’s lesson remains: verification starts with simple expectations and edge cases. For an AI-enabled simulator, ask:

  • Did the code implement the model?
  • Does the model represent the relevant mechanism?
  • Can its parameters be identified and intervention correspond to a real action?
  • How does uncertainty reach the recommended decision, and what changes when new data disagree?

Ten thousand runs cannot rescue a wrong mechanism. Neither can a foundation model. The opportunity is to build constrained experimental worlds where interventions can be tested virtually, uncertainty stays visible, and real-world feedback keeps the model honest.

Sources