Forecasting the local progression of the Covid-19 epidemic from medical emergency calls: the example of the Paris area

Stéphane Gaubert, Marianne Akian, Xavier Allamigeon, Marin Boyet, Baptiste Colin, Théotime Grohens, Laurent Massoulié, David P. Parsons, Frédéric Adnet, Érick Chanzy, Laurent Goix, Frédéric Lapostolle, Éric Lecarpentier, Christophe Leroy, Thomas Loeb, Jean-Sébastien Marx, Caroline Télion, Laurent Tréluyer, Pierre CarliarXiv 2020 · DOI · OpenReview

Abstract: We portray the evolution of the Covid-19 epidemic during the crisis of March-April 2020 in the Paris area, by analyzing the medical emergency calls received by the EMS of the four central departments of this area (Centre 15 of SAMU 75, 92, 93 and 94). Our study reveals strong dissimilarities between these departments. We show that the logarithm of each epidemic observable can be approximated by a piecewise linear function of time. This allows us to distinguish the different phases of the epidemic, and to identify the delay between sanitary measures and their influence on the load of EMS. This also leads to an algorithm, allowing one to detect epidemic resurgences. We rely on a transport PDE epidemiological model, and we use methods from Perron-Frobenius theory and tropical geometry.

Introduction

During the initial wave of COVID-19 in early 2020, public health officials faced a significant challenge: "lagging" data. Hospitalization and death counts reflect infections that occurred weeks prior, making it difficult to assess the immediate impact of policy changes like lockdowns. Standard epidemiological models, such as SEIR ODEs, often fail to account for precise incubation delays and spatial granularity. This research addresses the need for a spatially granular, early-warning system capable of detecting epidemic trends in real-time, rather than waiting for ICU capacity to be reached.

Data Overview

The study utilizes high-resolution datasets from the Paris area (departments 75, 92, 93, and 94) between February and May 2020. The primary data source is the SAMU (Center 15) records, encompassing over 170,000 patient files tagged for COVID-19 suspicion. These records are categorized by outcome: medical advice, ambulance dispatch, or Mobile ICU (MICU) transport. This data is validated against hospital-level data from Santé Publique France (SPF), including Emergency Room admissions and ICU entries.

Flowchart of SAMU call processing and hospital admission outcomes.

Methodology

The researchers combined clinical data science with sophisticated mathematical frameworks to model the epidemic's progression:

  • Transport PDEs: Unlike standard compartmental models that assume instantaneous transitions, the authors use Partial Differential Equations to model the "age" of an infection, accurately capturing the ~5-day incubation period.
  • Tropical Geometry & "Log Glasses": Recognizing that early-stage epidemic growth is exponential, the authors analyzed the logarithm of the data. Using "tropical" methods, they approximated these log-curves as piecewise linear functions. Each "break" in the line corresponds to a change in sanitary policy, such as the start of a lockdown.
  • Robust Optimization: To handle noisy call data, the model employed the $\ell_1$-norm (minimizing absolute errors) instead of standard least squares. This approach reduces sensitivity to "outliers," such as sudden spikes in calls triggered by media announcements rather than actual infections.

Key Insights

The analysis revealed several critical patterns in the epidemic's behavior:

  • Calls as a Leading Indicator: A consistent 6–7 day lead time was observed between the peak in emergency calls and the subsequent peak in ambulance dispatches and hospital load. This window is crucial for proactive resource allocation.
  • The "Announcement Effect": The peak for "medical advice" calls occurred on March 13th, prior to the lockdown. This spike was driven by public anxiety following government announcements rather than the biological peak of the virus. Consequently, "ambulance dispatch" proved to be a more reliable metric for actual disease progression.
  • Spatial Heterogeneity: The epidemic moved at different speeds across departments. For instance, the doubling time in Seine-Saint-Denis (93) was 4.2 days, compared to 5.9 days in central Paris (75), likely due to differences in population density and transit reliance.
Daily activity of SAMU on a log scale, showing the lead time of calls over dispatches and the "announcement effect."

Results and Conclusion

The piecewise linear approximation of the log-data provided a remarkably tight fit, confirming that epidemic phases are mathematically distinct and tied to specific policy interventions. The model identified a 10–14 day delay between policy implementation (e.g., school closures) and its visible effect on severe case growth rates.

The resulting monitoring algorithm, which uses the slope of log-curves to trigger "doubling time" alarms, was operationally deployed across the four SAMU centers of the Paris Public Hospitals (AP-HP). While early signals like calls for advice are sensitive to news cycles, cross-referencing them with ambulance dispatch data provides a robust early-warning system for detecting regional resurgences. Future work aims to apply these methods at even finer spatial scales to identify micro-clusters before they escalate.

Summary updated

Abstract

We portray the evolution of the Covid-19 epidemic during the crisis of March-April 2020 in the Paris area, by analyzing the medical emergency calls received by the EMS of the four central departments of this area (Centre 15 of SAMU 75, 92, 93 and 94). Our study reveals strong dissimilarities between these departments. We show that the logarithm of each epidemic observable can be approximated by a piecewise linear function of time. This allows us to distinguish the different phases of the epidemic, and to identify the delay between sanitary measures and their influence on the load of EMS. This also leads to an algorithm, allowing one to detect epidemic resurgences. We rely on a transport PDE epidemiological model, and we use methods from Perron-Frobenius theory and tropical geometry.
BibTeX
@misc{gaubert2020forecasting,
  author = {Stéphane Gaubert and Marianne Akian and Xavier Allamigeon and Marin Boyet and Baptiste Colin and Théotime Grohens and Laurent Massoulié and David P. Parsons and Frédéric Adnet and Érick Chanzy and Laurent Goix and Frédéric Lapostolle and Éric Lecarpentier and Christophe Leroy and Thomas Loeb and Jean-Sébastien Marx and Caroline Télion and Laurent Tréluyer and Pierre Carli},
  title = {{Forecasting the local progression of the Covid-19 epidemic from medical emergency calls: the example of the Paris area}},
  year = {2020},
  eprint = {2005.14186},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2005.14186},
}
Figure