Conformalized Physics-Informed Weak Regression for Parameter Estimation in Epidemic Models

Epidemics are often described by mathematicians using a system of compartments, much like buckets of water connected by pipes. In these models, people move from being susceptible to being infected, then to recovering or dying. The speed at which they move between these buckets is determined by physical parameters: how contagious a virus is, how long an infection lasts, and how quickly people recover. Knowing these numbers is vital. They tell public health officials how fast an outbreak will grow, how many hospital beds might be needed, and whether a lockdown or a vaccine campaign will work. For years, scientists have tried to find these hidden numbers by looking at real-world data, such as daily counts of sick people. This is a difficult puzzle because the data is often messy, incomplete, or collected at irregular intervals. In recent years, a popular method called "physics-informed neural networks" has been used to solve this. These are complex computer programs that learn to guess the parameters by training for hours or even days, but they often provide no clear measure of how wrong their guesses might be.

A new study from the Technical University of Denmark offers a different, much faster approach that also tells scientists exactly how much they can trust the results. The researchers developed a method that treats the problem as a straightforward math puzzle rather than a learning exercise. Instead of trying to guess the speed of change in the data, which is notoriously noisy and hard to measure, their method looks at the total accumulation of cases over time. By integrating the data, they smooth out the noise and can solve for the parameters almost instantly. The team tested this new technique on eleven different disease models, ranging from a simple model of COVID-19 to complex models of Zika and Ebola. They found that their method was often hundreds or even thousands of times more accurate than the popular neural network approaches, and it produced answers in milliseconds rather than hours. Crucially, they wrapped this fast calculation in a statistical safety net that generates a range of likely values for every parameter. This range is narrow when the data is clear and wide when the data is too messy to give a precise answer, giving scientists a honest picture of their uncertainty.

The core of this new method is a shift in how the data is read. Traditional methods often try to calculate the rate of change at every single moment, like measuring the speed of a car by looking at how far it moved in a split second. If the data has any noise or gaps, this speed calculation becomes wildly inaccurate. The Danish team realized that they could instead look at the total distance traveled over a longer period. In the context of an epidemic, this means looking at the total number of people who have been infected over a week or a month, rather than trying to pinpoint the exact moment a new infection occurred. This approach, which the author calls "weak regression," is far more robust against noise. It allows them to estimate the starting conditions of an outbreak and the transmission rates simultaneously, solving the entire system in a single, closed-form calculation. Because the math is linear and direct, it does not require the massive computational power or the long training times associated with neural networks.

To ensure these fast answers were reliable, the researchers added a layer of statistical calibration. In many scientific fields, it is difficult to know how accurate a model is because the true answer is unknown. The team solved this by using the disease models themselves to generate a massive library of fake epidemics with known answers. They ran their fast estimator on thousands of these simulated outbreaks, each with slightly different parameters and added noise, to see how often the method was right and how wide the error margins needed to be. This process created a "calibration population" that allowed them to attach a confidence interval to every real-world estimate. The result is a set of simultaneous intervals for all the parameters of a disease model at once. If the data is clear, the interval is tight; if the data is sparse or the dynamics are too complex to resolve, the interval widens automatically to reflect that uncertainty. This ensures that the scientists are never misled by a precise-looking number that is actually wrong.

When the team applied this method to eleven standard benchmarks used to test disease models, the results were striking. In fifteen out of twenty-two different scenarios involving various diseases and noise levels, their method produced the most accurate estimates. For models where the data was dense enough to resolve the dynamics, their estimates were often one to four orders of magnitude more accurate than those reported by the neural network methods. In terms of speed, the difference was even more dramatic. While the neural networks required training times ranging from twenty minutes to over ninety hours, the new method produced a full set of estimates and their uncertainty intervals in less than a millisecond. Even when the researchers included the time needed to generate the calibration data and verify the results, the entire process took less than two minutes. This speed means that public health officials could potentially update their understanding of an outbreak in real-time as new data comes in, rather than waiting days for a model to converge.

The study also tested the method on real historical data, specifically the 1978 influenza outbreak in a British boarding school. Here, the researchers compared their results against a detailed Bayesian analysis that had been published previously. Their method produced intervals that overlapped with the credible intervals of the Bayesian study, confirming that the fast, simple approach could capture the same biological reality as the more complex, computationally heavy methods. In this case, they also showed how to incorporate prior knowledge, such as a known recovery rate, to further tighten the estimates. The method proved flexible enough to handle time-varying transmission rates, where the contagiousness of a virus changes over time due to interventions or new variants. In these tests, the new approach again outperformed neural network baselines, providing accurate estimates of how transmission changed over time in a fraction of the time.

The researchers emphasize that this method does not replace other forms of analysis but rather complements them. It is particularly powerful when the data is sufficient to resolve the underlying dynamics of the disease. However, the method also knows its limits. When the data is too sparse or the noise is too high, the confidence intervals naturally expand to cover a wide range of possibilities, signaling that the data simply cannot support a precise answer. This honesty about uncertainty is a key feature. The study shows that for a wide class of epidemic models, the complex, black-box training of neural networks is not necessary. By returning to the fundamental physics of the problem and using a robust, integrated approach, scientists can achieve higher accuracy and better uncertainty quantification in a matter of milliseconds. This opens the door to more responsive and reliable modeling for future outbreaks, where speed and clarity are just as important as the numbers themselves.