By Vicki Glaser
June 22, 2009 | In the early 1990’s the co-founders of Archimedes, physician-mathematician David Eddy and physicist Len Schlessinger, began building the Archimedes Model (AM) as a consulting project for Kaiser Permanente (KP). The AM is a mathematical model that simulates human physiology (and the surrounding health care system) by creating virtual trial populations. Each individual does not represent a real person; rather the virtual people, in aggregate, represent a target population. Drawing on public datasets and clinical trials, the model can accurately simulate and predict the outcomes of trials.
“What sets Archimedes apart from other modeling companies is that not only does the Archimedes Model capture and integrate actual physiology into a single model, it also incorporates the interaction of individuals with the greater health care system. This allows Archimedes’ clients to understand the direct effects of interventions on actual health outcomes, such as in clinical trials, and the impact of these interventions in real world settings,” says Peter Alperin, medical director at Archimedes.
The AM evolved over a decade under the KP umbrella. In 2003, a seminal paper in Diabetes Care (Eddy, D. et al. 26:3093; 2003) validated the model in diabetes. In that study, funded by Bristol-Myers Squibb, Archimedes modeled 18 published trials selected by the American Diabetes Association (ADA), demonstrating its ability to predict the trial outcomes retrospectively.
Recognizing the model’s commercial potential, KP spun out Archimedes as an independent, for-profit entity in January 2006. “That enables us to work with clients that might not want to work under the KP umbrella, such as large pharma companies or [health insurers like] Humana,” says Mika Newton, Archimedes’ director of sales.
A promising real-world application of the AM is to study the safety of diabetes drugs. “The FDA has new requirements to show that a diabetic drug does not have cardiovascular [CV] side effects,” explains Newton. “To do that, you need enough CV events in the control arm to show that there is no impact on [event rate] by the new drug.” The AM model can estimate the underlying event rate in a control population and demonstrate how that rate would change by modifying population characteristics or inclusion/exclusion criteria.
The Collaborative Atorvastatin Diabetes Study (CARDS), sponsored by Pfizer and reported in 2004, evaluated the effects of atorvastatin in type 2 diabetes patients, focusing on CV outcomes. Before the trial ended, the investigators invited Archimedes to model the trial prospectively. Archimedes created a virtual population reflecting the characteristics of the trial population, modeled the effects of atorvastatin, and ran the model for five ‘virtual’ years. The results, revealed shortly before presentation of the actual trial data, closely matched the trial results.
The AM relies on publicly available datasets and clinical trial data that form the basis for the relational differential equations. Archimedes uses new trial results to challenge the model and determine if the existing equations adequately represent these new data. In this way, the model evolves to incorporate new knowledge about disease pathways, interventions, and outcomes.
“When we recruit virtual patients for our trials, they are individual virtual people that we build by pulling them out of a publicly available dataset,” explains Alperin. He says it’s easier to teach mathematically oriented modelers that create the algorithms a little medicine than teaching medics heavy math.
The equations are solved simultaneously using a technique called event queuing. This asks, “What is the next thing that will happen?” Alperin explains. “Each time there’s an event, the output of the model is the health outcome that took place. Events are not necessarily outcomes—they can be something as massive as an MI or as simple as going to see the doctor. After every event, all the equations are recalculated,” resulting in a reordering of the predictions of future events.
With a grant from the American Cancer Society (ACS), Archimedes built cancer into the AM with the goal of developing a simple, prioritized standard for a preventive encounter that pinpoints lifestyle factors and specifies tests for physiologic parameters. The model would predict benefit -- avoiding premature death, extending lifespan—while also simulating various kinds of health care systems to achieve the greatest benefit with acceptable economic metrics (Kahn, R. Circulation 118:576-585; 2008).
Robert Smith, ACS’ director of cancer screening, says that the trials themselves could not be done—they would be too enormous, too expensive, and too lengthy. The ACS plans to use the AM to predict the outcomes of different cancer screening interventions, such as relative costs and benefits of increasing adherence with regular testing vs. improvements in quality, and the combined effects of each. Smith sees clear advantages with AM. “You can simulate unique populations or organizations and say, ‘There is an opportunity to deliver preventive health and here is what the impact would be on your employees and your bottom line.’”
Another partner is GE Healthcare, which linked up with Archimedes to focus on heart disease and to explore the cost effectiveness of screening for MI risk among millions of asymptomatic people. The AM can simulate scenarios for which there is little existing data and assimilate new information. This enables testing of the effects of a newly discovered biomarker, for example, without having to rebuild the model to accommodate a new component.
GE and Archimedes ran a simulated clinical trial in a virtual population over 30 years to predict outcomes for various screening and preventive strategies. Results showed that unconditional treatment with statins would be the most cost effective approach, but if imaging of obstructed coronary arteries could boost compliance rates, then imaging would be the most cost effective screening tool.
Back to the Future
Archimedes has also developed models of kidney function, metabolic syndrome, and obesity. The obesity model can be applied to study the effects of various weight reduction methods and how those affect biomarkers and use of health care resources. “We have a laundry list of what we would like to do next,” says Newton, including mental health (schizophrenia and bipolar disorder), inflammation, Parkinson’s, and multiple sclerosis.
The AM runs on a dedicated grid environment, performed as fee-for-service consulting. “The output of the model is very complex and it takes our scientists and medical team to interpret it appropriately,” says Newton. But that business model is changing following a $15.6-million grant from the Robert Wood Johnson Foundation to build the Archimedes Healthcare Simulator (ARCHeS) web-based interfaces. That work would provide access to federal, state, and local governments, for example, to predict future needs. Archimedes would provide the model at cost to subscribing organizations. Newton anticipates the first software releases in about two years.