Building Archimedes - a Q&A with 'Dr. Data' - ' The Archimedes Difference '

The Archimedes Difference

What makes this model better than and different from others out there?

The closest comparison to Archimedes in clinical medicine (as opposed to biological models or operations research models) is a Markov model, named after the Russian mathematician who developed it at in the early 1900s. Markov models are fundamentally different than Archimedes. A Markov model breaks a disease into a small number of discrete clinical "states" and breaks time into discrete intervals, usually a year. At any time, a patient will be in one and only one of the states, and can move from one state to another at the end of each time interval. The progression of a disease is determined by the probabilities that a patient will move from one state to another (called "transition probabilities"). For example, diabetic nephropathy has been represented in several Markov models as four states: at any time a patient can either "have microalbuminuria," "have proteinuria," "have end-stage renal disease (ESRD)," or "be dead," with patients being able to move from one state to the next at annual intervals. To draw on an answer to an earlier question, cardiovascular complications of diabetes have been modeled by saying that a patient can be in one of two states: "has no cardiovascular complications" or "has cardiovascular complications." The effects of treatments are modeled as changes in the transition probabilities. Thus the most obvious difference is that Markov models are very superficial; there is no attempt to model the underlying physiology or biological variables, or to model the effect of a treatment in a biologically realistic way. A second obvious difference is that they handle time very crudely. Patients hop from state to state at discrete time intervals, usually one year. Whatever might happen to a patient in the intervals between each yearly hop is missed. I can say these things without intending any disrespect, because all my previous models were Markov models. In fact, I won an international prize in mathematics for my contributions to the theory and application of Markov models. They have valuable uses; they are just not up to the task of what we are trying to do.

What are the other features that make Archimedes different from other models in healthcare?

Archimedes is distinguished from other models in healthcare by several features. First, it is a full simulation. All the objects that are pertinent to a real problem (as determined by clinicians and administrators) are included in the model, one-to-one, and are connected in an organic and continuous way that represents how they interact in reality. Second, it spans a broad range of diseases and interventions, enabling the types of comparisons required for things such as setting priorities. Third, it goes from soup to nuts – spanning from biological details to the logistics and structures of health-care systems. Fourth, it is written at a deep level of detail. The physiology models include anatomy and biological variables at the level of detail that clinicians require, guidelines and disease management programs are designed, and clinical research is conducted. The models of care processes and system resources are written at the level of detail of administrative decisions, protocols and continuous quality improvement projects. Fifth, it is continuous in time. All the variables are continuously interacting and any event can occur at any time. The times between events can be as short as minutes or as long as years. Sixth, biological variables that are continuous in reality are represented continuously in the model. They are not broken into discreet states, stages or strata, although they can be translated into these artifacts when clinical taxonomies call for that. Taken together, these characteristics make the model "natural"; its structure, detail and operation correspond to how we think about clinical and administrative problems. This is what enables the model to address such a wide variety of problems, including ones that were not contemplated when the model was built.

A final – and we believe critical – distinction is that because the model is so realistic, we can validate it by comparing what happens in the virtual world of the model with what happens in reality.

Does the model account for human behavior?

We do have patient behavior and physician behavior in the model. For example, if there is evidence that physicians vary in how they follow a guideline, or that patients have different thresholds for going to an emergency department for chest pain, we include it.