DESA: An R Package for Detecting Epidemics using a School-Absenteeism Surveillance Framework

This paper introduces DESA, an R package available on CRAN that models, detects, and evaluates alerts for influenza epidemics using elementary school absenteeism data, while also providing simulation tools for community-level research.

The Gist

Imagine you are the captain of a massive ship (a city or region) trying to navigate through a foggy sea. Your biggest fear is a sudden, invisible storm (an epidemic like the flu) that could damage your crew (the population). You can't see the storm coming until the waves start crashing over the deck, but by then, it's often too late to take evasive action.

This paper introduces DESA, a new digital "radar system" built by a team of researchers to help public health officials spot these storms early. Instead of waiting for people to go to the hospital (which is slow and often happens too late), DESA listens to a different, quieter signal: school absences.

Here is how the system works, explained through simple analogies:

1. The Early Warning Signal: The "Canary in the Coal Mine"

Think of elementary school children as the "canaries in the coal mine." They are often the first to catch a cold or the flu and the first to stay home.

• The Old Way: Waiting for doctors to report sick patients is like waiting for the ship's hull to crack before you know there's a storm.
• The DESA Way: It watches the "attendance board." If suddenly, 10% of kids in a school are missing, DESA treats that like a flashing red light on the radar, suggesting a storm is brewing before anyone even knows they are sick.

2. The Engine: A "Weather Simulator"

To make sure their radar works, the researchers needed to test it. But you can't wait for a real flu pandemic to happen just to test a tool.

• The Analogy: Imagine a video game where you can simulate a hurricane. You can tweak the wind speed and rain, then see if your storm shelter holds up.
• How DESA does it: The package includes a powerful simulator. It builds a fake population (families, houses, schools) and then "infects" it with a virtual virus. It watches how the fake virus spreads and how many fake kids stay home. This allows scientists to test their detection tools thousands of times without anyone actually getting sick.

3. The Brain: The "Detective's Logic"

DESA uses a special mathematical formula (a "lag-logistic regression") to connect the dots.

• The Analogy: Think of a detective trying to solve a mystery. The detective knows that the "crime" (the virus) happened a few days before the "clue" (the child staying home) appeared.
• How it works: The system looks at yesterday's absence numbers to predict today's risk. It accounts for the fact that flu season happens every winter (seasonal patterns) and that some years are worse than others. It asks: "Is this spike in absences just a normal cold, or is it the start of a massive outbreak?"

4. The Scorecard: "Did the Alarm Ring at the Right Time?"

When the system raises an alert, how do we know if it was good? The paper introduces a "Scorecard" with three main ways to grade the alarm:

• The "False Alarm" Rate (FAR): Did the system scream "Wolf!" when there was no wolf? (We don't want to panic the town for no reason).
• The "Delay" Score (ADD): If there was a wolf, how many days late was the alarm? (An alarm that rings after the wolf has already eaten the sheep is useless).
• The "Timing Quality" (ATQ): This is the most sophisticated score. Imagine the perfect time to ring the alarm is exactly 14 days before the storm hits.
  • Ringing it 14 days early? Perfect score.
  • Ringing it 1 day early? Good score.
  • Ringing it 20 days early? Too early, might cause panic.
  • Ringing it 1 day late? Too late, people are already wet.
  • The system tries to find the "Goldilocks" zone—not too early, not too late.

5. Why This Matters

The paper explains that DESA is a free tool (like a free app) that anyone can download. It helps public health officials:

• Plan better: If they know a flu wave is coming in two weeks, they can send out vaccines or antiviral medicine before the hospitals get overwhelmed.
• Save money: It's cheaper to stop a small outbreak than to fight a massive one.
• Test ideas: They can run simulations to see, "What if we close schools earlier?" or "What if we have a different virus?" without risking real lives.

In a Nutshell

DESA is a smart, free computer program that acts like a crystal ball for flu season. By watching how many kids stay home from school, it predicts when a flu epidemic is about to hit, giving communities a head start to protect themselves. It's like having a weather forecast that tells you it's going to rain before the first cloud appears.

Technical Summary

Here is a detailed technical summary of the paper "DESA: An R Package for Detecting Epidemics Using a School-Absenteeism Surveillance Framework."

1. Problem Statement

Traditional influenza surveillance systems rely on clinical reports and laboratory-confirmed cases. These methods often suffer from significant delays and may fail to capture the real-time dynamics of an epidemic, especially when a large portion of infected individuals do not seek medical care.

• The Gap: There is a need for early warning systems that utilize alternative data sources, such as school absenteeism records, which can provide timely indicators of community-wide epidemic arrivals.
• The Challenge: Previous research (specifically Vanderkruk et al.) demonstrated the efficacy of school absenteeism surveillance but relied on computationally intensive individual-level models (ILMs) that were prohibitive for simulating large populations (e.g., cities or regional municipalities). Furthermore, there was a lack of accessible, generalized software tools for researchers to implement these models, evaluate alert metrics, and simulate scenarios without real-world outbreak data.

2. Methodology

The paper introduces DESA (Detecting Epidemics from School Absenteeism), an R package designed to model epidemics, raise alerts, evaluate performance, and simulate data. The methodology is built on four core pillars:

A. Epidemic Detection Model

DESA utilizes a regional-level seasonal mixed-effects binary logistic regression model to predict the probability of at least one laboratory-confirmed influenza case ($\theta_{tj}$) based on school absenteeism.

• Model Structure: A lag-logistic regression model incorporating:
  • Lagged Absenteeism: Mean daily absenteeism percentages with lag times ($0$to$l$).
  • Seasonality: Trigonometric functions (sine and cosine) to account for annual influenza cycles.
  • Random Effects: A random effect term ($\gamma_j$) for each school year to address intra-correlation and heterogeneity across years.
• Alert Mechanism: An alert is triggered on day $t$ if the predicted probability $\theta_{tj}$ exceeds a user-defined threshold ($\theta^*$).

B. Evaluation Metrics

To assess the timeliness and accuracy of alerts, DESA implements several metrics:

• False Alarm Rate (FAR): The proportion of alerts raised outside the "true alarm window" (defined as 14 days prior to the reference date).
• Accumulated Days Delayed (ADD): Measures the delay between the optimal alert time and the actual first alert.
• Alert Time Quality (ATQ): A continuous penalty function that penalizes early alerts more heavily than late ones, allowing for a nuanced assessment of timing.
• Aggregate Metrics: Average ATQ (AATQ), First ATQ (FATQ), and their weighted variants (WAATQ, WFATQ) which account for the increasing amount of training data available over time.

C. Simulation Framework

To overcome the computational limitations of previous individual-level models, DESA implements a Stochastic SIR (sSIR) framework.

• Population Generation: A hierarchical approach simulates catchment areas, schools, and households (with and without children), assigning spatial coordinates and demographic attributes.
• Disease Dynamics: Instead of tracking every individual pair (as in ILMs), the sSIR model simulates transitions between Susceptible, Infected, and Removed compartments at the population level using binomial distributions. This allows for efficient simulation of large populations (e.g., ~200,000 individuals).
• Absenteeism Synthesis: The package generates synthetic absenteeism data by combining:
  • Baseline Absenteeism: A fixed probability (e.g., 5%) of absence due to non-influenza causes.
  • Infection-Related Absenteeism: A high probability (e.g., 95%) of absence for infected students during their infectious period.
  • Reporting Delays: Simulated reporting lags and confirmation rates to mimic real-world surveillance noise.

3. Key Contributions

1. Computational Efficiency: By replacing individual-level models with a stochastic SIR approach, DESA enables the simulation of epidemics in large populations (cities/regions) that were previously computationally infeasible.
2. Generalization: While previous work was specific to the Wellington–Dufferin–Guelph Public Health Unit, DESA generalizes the methodology, allowing application to diverse regions and populations.
3. Comprehensive Toolset: The package provides a complete workflow from data generation (population, epidemic, absenteeism) to model fitting, alert generation, and multi-metric performance evaluation.
4. Accessibility: The package is freely available on CRAN and GitHub, lowering the barrier to entry for public health officials and researchers to develop and test early warning systems.
5. Flexible Evaluation: It offers a suite of metrics (FAR, ADD, ATQ variants) allowing users to optimize models based on specific surveillance goals (e.g., prioritizing early detection vs. minimizing false alarms).

4. Results

The paper demonstrates the package's capabilities through a reproducible workflow using simulated data:

• Simulation: The authors simulated a population of ~199,669 individuals across 16 catchment areas and generated 10 epidemic replicates.
• Model Performance: Using the eval_metrics function, the package identified optimal model parameters (lag and threshold) for different metrics.
  • Example Findings: The optimal lag varied by metric (ranging from 1 to 5 days), and optimal thresholds ranged from 0.1 to 0.45.
  • Visualization: The package successfully generated plots comparing absenteeism rates, confirmed cases, and alert timings against reference dates, demonstrating the system's ability to provide early warnings before the peak of the epidemic.
• Validation: The package passed all CRAN checks (R CMD check) without errors or warnings, and unit tests were implemented using the testthat framework.

5. Significance

• Public Health Impact: DESA provides a practical tool for public health authorities to improve the timeliness of interventions (e.g., vaccination campaigns, antiviral distribution) by detecting influenza outbreaks earlier than traditional clinical surveillance.
• Research Utility: It enables researchers to test surveillance strategies and model parameters in a controlled, simulated environment without waiting for real-world outbreaks. This is particularly valuable given the disruption of school data collection during the COVID-19 pandemic.
• Future Directions: The modular design allows for the adaptation of the framework to other infectious diseases beyond influenza and supports the development of more robust, data-driven early warning systems.

In summary, DESA represents a significant advancement in epidemic surveillance software, bridging the gap between complex statistical modeling and practical, accessible application for public health decision-making.