Traditional clinical research often treats patient outcomes as binary events: a patient either survives or dies, or they either respond to a treatment or they do not. However, real-world health is a continuous journey through various stages of severity, recovery, and complication. To capture this complexity, researchers have converged on multistate modeling, a framework that treats patient health as a sequence of transitions between discrete, well-defined states.
By modeling the "middle" of the patient journey, these methods allow clinicians and policymakers to project long-term outcomes from short-term data, account for competing risks, and evaluate the true economic value of medical interventions.
From Snapshots to Trajectories
The core intuition behind this research direction is that health risk is an evolving signal rather than a static attribute. As Jenna Wiens, John V. Guttag, and Eric Horvitz argue in Patient Risk Stratification for Hospital-Associated C. diff as a Time-Series Classification Task, clinical risk follows a progression. Vulnerabilities, such as those caused by antibiotic exposure, create sustained high-risk states that do not simply revert.
To model this, researchers use State-Transition Models (STMs), often implemented as Markov models. In this setup, a patient cohort is distributed across various health states—such as "Healthy," "Progressed Disease," or "Post-Complication"—and moves between them over fixed time intervals (cycles) based on transition probabilities.
This approach solves a fundamental problem in clinical trials: trials are usually too short to observe lifetime impacts. As Léon Nshimyumukiza and colleagues demonstrate in their study on stroke prevention, Markov modeling allows researchers to take short-term trial data and extrapolate it to project lifetime health outcomes and costs.
Capturing Granular Clinical Progressions
While simple models might only use three states (the "Illness-Death" model), recent work has increased granularity to better reflect hospital realities. For instance, during the COVID-19 pandemic, Michael Roimi, Rom Gutman, and others developed multistate survival models that tracked transitions between "Moderate," "Severe," and "Critical" care.
This granularity is essential for two reasons:
- Competing Risks: A patient might be discharged (recovery) or die (competing risk). Traditional models often struggle to handle these simultaneous possibilities, but multistate models calculate transition-specific hazards for each path.
- Resource Planning: By understanding the probability of moving from a general ward to the ICU, hospital planners can forecast bed demand more accurately than by using simple average lengths of stay. Hagai Rossman and Jonathan Somer used this logic to show how hospital load directly impacts mortality rates by altering these transition dynamics.
Quantifying Value through Simulation
A major application of this framework is in Cost-Utility Analysis. By assigning a "utility" weight (representing quality of life) and a cost to each state, researchers can calculate Quality-Adjusted Life Years (QALYs).
In Cost-Effectiveness of Pembrolizumab vs. Standard-of-Care Chemotherapy, Min Huang and James M. Pellissier utilize partitioned-survival modeling—a variant of multistate modeling—to track how much time patients spend in "Progression-Free" versus "Progressive Disease" states. The area between these survival curves represents the survival benefit, which is then weighed against the high cost of immunotherapy.
Similarly, Ivandito Kuntjoro and Edgar Tay employ a two-phase framework for heart valve implants: a decision tree captures the immediate, high-resolution risks of surgery, while a Markov model tracks the long-term chronic health states over the patient's remaining life.
Correcting for Bias in Real-World Data
When moving from controlled trials to real-world Electronic Health Records (EHR), multistate models become a tool for causal inference. Nir Keret highlights that retrospective data often suffers from "survival bias"—we only see the patients who lived long enough to be recorded. By using an illness-death stochastic model, researchers can account for this bias, unlocking the power of "prevalent" information that was previously discarded.
This hybrid approach—combining observational data with validated simulation—is now being used by researchers like Dejana Braithwaite and Rafael Meza to personalize cancer screening. By simulating how chronic conditions like COPD alter the transition probabilities of lung cancer, they can estimate whether the benefits of screening outweigh the risks of overdiagnosis for a specific individual.
The Future: Deep Learning and Microsimulation
The field is currently evolving in two directions. First, researchers like Philipp Kopper and David Rügamer are integrating deep learning with these models. Their DeepPiecewise Exponential Models allow neural networks to process unstructured data (like medical images) to predict the transition probabilities between states.
Second, there is a shift from cohort-level Markov models to individual-level microsimulations. As seen in Stewart B. Harris's work on diabetes management, microsimulation tracks "virtual patients" one by one. This allows the model to remember a patient's unique history—such as how many previous hypoglycemic events they've had—which influences their future transition probabilities, leading to far more precise and personalized projections of health.