🔬 physics

A Universal Convolution-Based Pre-processor to Correct the Prevalence-Incidence Gap in SIR, SEIR, and SIRS Modeling

This paper proposes a universal convolution-based pre-processor that corrects the fundamental methodological error of calibrating prevalence-based compartmental models (SIR, SEIR, SIRS) with raw incidence data, thereby eliminating systematic predictive biases and bridging the gap between clinical reporting and mechanistic epidemic forecasting.

Original authors: Jose de Jesus Bernal-Alvarado, David Delepine

Published 2026-02-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jose de Jesus Bernal-Alvarado, David Delepine

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Problem: Mixing Up "New Arrivals" with "People Inside"

Imagine you are trying to predict how crowded a busy concert hall will get. You have two types of data:

Incidence (New Arrivals): The number of people walking through the front door right now.
Prevalence (The Crowd): The total number of people currently standing inside the hall.

The paper argues that for years, scientists have been making a massive mistake. They are taking the number of new people walking in (Incidence) and pretending it is the total number of people inside (Prevalence) to feed into their computer models.

The authors call this a "fundamental methodological error." It's like trying to guess how full a bathtub is by only looking at how fast the water is pouring in from the faucet, while ignoring that the drain is open and water is leaving at the same time.

Why This Matters: The "Peak" Mistake

When scientists use this wrong data, their models fail in two specific ways:

They get the timing wrong: The model thinks the crowd will peak (be most crowded) too early.
They get the size wrong: The model thinks the crowd will be much smaller than it actually is (underestimating the peak by up to 50%).

The paper shows that simply making the model more complicated—by adding more "rooms" to the building (like adding a waiting room for people who are sick but not contagious yet)—doesn't fix the problem. If you feed the wrong data (new arrivals) into a complex building, the whole structure collapses. The error just spreads to the new rooms.

The Solution: A "Universal Pre-Processor"

The authors propose a simple fix: a "pre-processor." Think of this as a special filter or a translator that you must use before you put any data into your model.

Instead of just counting the people walking in today, this filter looks at:

How long people stay: If a sick person stays for 5 days, the filter remembers that the people who walked in 4 days ago are still inside.
How fast they leave: It accounts for the fact that people recover and leave the "sick" room every day.
Missing data: It acknowledges that not every sick person gets reported (some stay home). It adds a "correction factor" to guess the real number based on the reported numbers.

The Analogy of the Leaky Bucket:
Imagine the sick population is a bucket.

Incidence is the water pouring in.
Recovery is a hole in the bottom letting water out.
Prevalence is the actual water level in the bucket.

The old way of doing things was to say, "The water level is exactly the same as the water pouring in right now." This is wrong because the hole in the bottom changes the level.

The new method uses a mathematical recipe (a convolution) that looks at all the water that poured in over the last few days, subtracts the water that leaked out, and gives you the true water level.

The "Magic Formula"

The paper suggests a specific math formula to do this translation. In plain English, it says:

"To find out how many people are sick today, don't just look at today's new cases. Look at all the new cases from the past few days, but give less weight to the ones from a long time ago (because those people have likely already recovered)."

They also add a "reporting rate" to account for the fact that many sick people aren't getting tested or reported, ensuring the model doesn't underestimate the true size of the outbreak.

The Bottom Line

The paper concludes that no matter how fancy your epidemic model is (whether it's a simple SIR model or a complex one with immunity fading), you cannot skip this step.

If you want your predictions to be accurate, you must first translate "new daily cases" into "active sick people" using this specific filter. Without it, even the most advanced models will be looking at the wrong picture, leading to bad predictions about when the outbreak will peak and how big it will get.

Technical Summary: A Universal Convolution-Based Pre-processor to Correct the Prevalence-Incidence Gap in SIR, SEIR, and SIRS Modeling

Problem Statement
Despite the widespread adoption of compartmental models (SIR, SEIR, SIRS) for epidemic forecasting since 2019, these frameworks consistently fail to align with real-world data validation, often resulting in peak underestimations of up to 50%. The authors identify a persistent "fundamental methodological error" at the core of this failure: the direct calibration of prevalence-based (stock) models using raw daily incidence (flow) data.

In standard epidemiological modeling, the infectious compartment $I(t)$ represents a prevalence variable (the stock of currently active cases). However, clinical reporting systems (e.g., WHO) typically provide incidence data (New Daily Cases, or NDC), which represents the flow of new infections. The paper argues that equating these two distinct mathematical entities—treating the time derivative of cumulative cases as the active case count—creates a structural misalignment. This error is not resolved by increasing model complexity; adding compartments for latency (SEIR) or waning immunity (SIRS) merely propagates the incidence-prevalence gap into the transition rates of the new compartments, leading to systematic biases in estimating latent periods and the "heavy tail" of endemicity.

Methodology
The authors propose a universal pre-processor protocol based on an exponentially weighted convolution to reconstruct active cases ( $I(t)$ ) from reported incidence ($NDC(t)$). The methodology proceeds in two stages:

Peak Position Correction: Initially, the authors address the temporal misalignment. Since prevalence is an accumulation of recent incidence minus recoveries, they propose integrating incidence over a sliding window defined by the average infectious period ( $1/\gamma$ ). This shifts the timing of the peak to align with the true prevalence curve.
Peak Amplitude and Structural Correction: Recognizing that a simple sliding window treats recovery as a "step function" (instantaneous recovery after exactly $1/\gamma$ days), the authors introduce a more rigorous mathematical correction. They model recovery as an exponential decay process and account for data under-reporting (asymptomatic cases and testing delays) via an ascertainment rate ( $p$ ).

The proposed universal transformation is defined as:
$I(t) \simeq \frac{1}{p} \int_{0}^{t} NDC(\tau)e^{-\gamma(t-\tau)} d\tau$

Where:

$NDC(\tau)$ is the observed daily incidence.
$\gamma$ is the recovery rate (defining the infectious period).
$e^{-\gamma(t-\tau)}$ acts as a survival probability weight, ensuring past incidence is discounted based on the probability that those individuals remain infectious at time $t$ .
$p$ is the ascertainment rate ( $0 < p \leq 1$ ), correcting for the fraction of true infections captured by the healthcare system.

Key Contributions

Identification of a Systematic Error: The paper formally categorizes the direct use of raw incidence data in stock-based models as a fundamental methodological error that persists across SIR, SEIR, and SIRS frameworks.
Universal Pre-processor Protocol: The authors introduce a specific convolution formula that serves as a mandatory prerequisite for calibrating any compartmental model. This protocol bridges the gap between clinical reporting (incidence) and mechanistic modeling (prevalence).
Correction of Complex Models: The study demonstrates that structural additions like latency (SEIR) or waning immunity (SIRS) do not inherently fix data misalignment. Without the proposed pre-processor, these advanced models inherit the same systematic biases, leading to incorrect estimates of transition rates and immunity loss.

Results
Simulation results presented in the paper indicate that without this transformation:

Models suffer from significant peak underestimations (up to 50%) and temporal misalignment (incidence peaks earlier than prevalence).
The "heavy tail" of endemicity is systematically underestimated.
In SEIR models, the latent period is incorrectly estimated because the infectious compartment is misaligned with the input data.
In SIRS models, the rate of immunity loss is misinterpreted due to the distortion of the curve's tail.

By applying the proposed convolution, the authors show that the reconstructed prevalence curve ( $I(t)$ ) successfully aligns with the theoretical requirements of the differential equations, correcting both the timing and the amplitude of the epidemic peak.

Significance and Claims
The paper claims that the proposed convolution transformation is not merely an optional refinement but a universal prerequisite for any compartmental framework. The authors argue that increasing structural complexity (moving from SIR to SEIR/SIRS) cannot compensate for the fundamental error of data misalignment. Instead, the "incidence-prevalence gap" must be closed at the data input stage.

The significance of this work lies in its assertion that the "fundamental methodological error" has been overlooked in existing literature, where convolution is often treated as a standalone proof rather than a universal calibration requirement. By implementing this pre-processor, the paper posits that epidemiological forecasting can achieve a level of accuracy that aligns clinical reporting with mechanistic modeling, thereby resolving the predictive failures observed in recent pandemic responses.