On May 26, New York Gov. Andrew Cuomo said what many people have been thinking about the COVID-19 model predictions: “They were all wrong.”
Indeed, on March 13, the Center for Disease Control estimated COVID hospitalizations in the United States could range from 2.4 to 21 million during the pandemic. As of Aug. 24, less than 60,000 have been hospitalized at any time. In light of the huge differences between the early projections and the actual results, one might conclude, as Cuomo did, that the computer models are useless as a tool for decision makers.
However, the models were not wrong; the assumptions were. The early model projections assumed the public’s risky behavior contributing to the spread of the disease would remain unchanged. But as a result of government-mandated social restrictions, people did change behavior, producing fewer infections. Had the modeling experts foreseen the behavioral changes, they could have input those assumptions with more accurate results.
There are two issues that modeling experts face. First, the mathematical equations must be correctly coded in the model. The debugging process to eliminate these errors is rigorous and precise. The second issue, determining the input assumptions, is often no better than educated guesses.
An example will help clarify this. Suppose you want to know how much extra food you need for house guests. Over time you record your expenditures when you have guests and you find that each guest adds about 4% per day to food costs. You place your mathematical model on a computer; input the number of guests and it estimates the food costs. You test your model with a few guests and find it accurate.
But one day some guests greatly exceed the projected costs. What happened? Your guests were teenage boys who eat twice as much as the adults upon which you based your model. Your model is accurate but used the wrong assumptions and produced an inaccurate estimate. To correct this problem, you update your model to account for the age of guests. Over time other model adjustments address new variables.
Similarly, COVID modeling experts, such as those on the N.M. Department of Health’s Modelling/Analytics Committee, have been adjusting both their models and their input assumptions with improving predictions. Input assumptions is one of the banes of computer modeling.
Early in the crisis full-scale lockdowns were the only options available to government officials to prevent hospitals from being inundated with virus victims. It worked. As a result of our governor’s assertive actions, the virus is coming under control.
But the result is one of the worst economic calamities in U.S. history, including a mushrooming national debt with interest payments that will encumber the country for decades. N.M.’s unemployment is over 12%, and many tax-paying businesses have permanently closed.
The lockdowns are analogous to killing all living plants in half of your garden to save it from an invasive weed. A more refined approach would instead identify and target the intruding species at an early stage, allowing the garden to continue to grow and produce.
Adaptive Resonance Theory, a form of Artificial Intelligence currently being studied at the University of New Mexico, can help. Unlike traditional statistical models, the basis of most COVID predictions, this form of AI is self-learning and self-organizing, similar to the processes used by all intelligent life. It is extraordinarily adept at identifying anomalies in data as they are collected from the field.
Had AI been integrated into the modeling process at an early stage, it could have helped experts to more quickly identify those factors, such as age and pre-existing conditions, that lead to hospitalizations and deaths, thus allowing targeted intervention and eliminating the need for economically damaging lockdowns.
As the state struggles with the question of how to reopen our communities, we urge state officials to employ the power of AI to help the simultaneous challenge of maintaining safety and economic stability.
