Estimating SARS-CoV-2 infection probabilities with serological data and a Bayesian mixture model

Benjamin Glémain, Xavier de Lamballerie, Marie Zins, Gianluca Severi, Mathilde Touvier, Jean‐François Deleuze, Fabrice Carrat, Pierre‐Yves Ancel, Marie-Aline Charles, Sofiane Kab, Adeline Renuy, Stéphane Le-Got, Céline Ribet, Mireille Pellicer, Emmanuel Wiernik, Marcel Goldberg, Fanny Artaud, Pascale Gerbouin-Rérolle, Mélody Enguix, C Laplanche, Roselyn Gomes-Rima, Lyan Hoang, Emmanuelle Correia, Alpha Amadou Barry, Nadège Senina, Julien Allègre, Fabien Szabo de Edelenyi, Nathalie Druesne‐Pecollo, Younes Esseddik, Serge Herçberg, Mélanie Deschasaux, Valérie Benhammou, Anass Ritmi, Laetitia Marchand‐Martin, Cécile Zaros, Elodie Lordmi, Adriana Candea, Sophie de Visme, Thierry Siméon, Xavier Thierry, Bertrand Geay, Marie‐Noëlle Dufourg, Karen Milcent, Delphine Rahib, Nathalie Lydié, Clovis Lusivika‐Nzinga, Grégory Pannetier, Nathanaël Lapidus, Isabelle Goderel, Céline Dorival, Jérôme Nicol, Olivier Robineau, Cindy J. Lai, Liza Belhadji, Hélène Espérou, Sandrine Couffin-Cadièrgues, Jean‐Marie Gagliolo, Hélène Blanché, Jean-Marc Sébaoun, Jean-Christophe Beaudoin, Laëtitia Gressin, Valérie Deldrève, Ouissam Ouili, Läétitia Ninove, Stéphane Priet, Paola Mariela Saba Villarroel, Toscane Fourié, Souand Mohamed Ali, Abdenour Amroun, Morgan Seston, Nazlı Ayhan, Boris PastorinoScientific 2024 · DOI

Abstract

The individual results of SARS-CoV-2 serological tests measured after the first pandemic wave of 2020 cannot be directly interpreted as a probability of having been infected. Plus, these results are usually returned as a binary or ternary variable, relying on predefined cut-offs. We propose a Bayesian mixture model to estimate individual infection probabilities, based on 81,797 continuous anti-spike IgG tests from Euroimmun collected in France after the first wave. This approach used serological results as a continuous variable, and was therefore not based on diagnostic cut-offs. Cumulative incidence, which is necessary to compute infection probabilities, was estimated according to age and administrative region. In France, we found that a "negative" or a "positive" test, as classified by the manufacturer, could correspond to a probability of infection as high as 61.8% or as low as 67.7%, respectively. "Indeterminate" tests encompassed probabilities of infection ranging from 10.8 to 96.6%. Our model estimated tailored individual probabilities of SARS-CoV-2 infection based on age, region, and serological result. It can be applied in other contexts, if estimates of cumulative incidence are available.

Introduction: The Limitations of Binary Serology

Standard SARS-CoV-2 antibody tests typically rely on fixed "positive/negative" cut-offs that discard nuanced data and ignore an individual's demographic risk. For instance, the Euroimmun ELISA test measures an Optical Density Ratio (ODR) where values < 0.8 are labeled "negative" and > 1.1 are "positive." This approach suffers from significant information loss near thresholds and fails to account for "pre-test probability"—the likelihood of infection based on context like age or location. Furthermore, a significant minority of infected individuals (5–24%) are "non-responders" who produce few antibodies and are almost always misclassified by standard cut-offs.

Methodology: A Bayesian Mixture Model

To address these flaws, researchers developed a Bayesian mixture model using the SAPRIS-SERO dataset, which contains 81,797 continuous anti-spike IgG test results from French adults collected after the first pandemic wave in mid-2020.

Instead of relying on manufacturer thresholds, the model treats the population as a weighted average of two distributions: $P(\text{ELISA} | \text{infected})$ and $P(\text{ELISA} | \text{uninfected})$. The uninfected group was modeled using a skew-normal distribution to capture the data's "tail," while the infected group used a mixture of two normal distributions to account for both antibody "responders" and "non-responders."

Probability density of ELISA results showing the mixture of infected and uninfected distributions.

Cumulative incidence (the weight of the infected group) was incorporated as a prior, using age and administrative region as covariates. To ensure model convergence, the researchers first calibrated the antibody distribution using a validation set of 319 RT-PCR-confirmed positive participants.

Personalized Infection Probabilities

The model's primary strength is its ability to provide personalized risk assessments. The results demonstrate that a "negative" result is not universally negative. For example, in the high-incidence Île-de-France (Paris) region, a 40-49-year-old with an ODR of 0.8—the very threshold for a negative result—actually has a 61.8% probability of past infection.

Variation in infection probability based on ELISA ODR, age, and geographic region.

The "indeterminate" zone (0.8–1.1) was found to be a wide spectrum, representing infection probabilities ranging from 10.8% to 96.6% depending on the individual's demographic profile. The model also estimated that 14.5% of infected individuals are non-responders, explaining why traditional sensitivity estimates (often cited at >90%) are frequently over-optimistic in real-world settings.

Key Findings and Epidemiological Impact

By refining the interpretation of serological data, the model provided more accurate epidemiological estimates for France after the first wave:

  • Cumulative Incidence: Estimated at 7.6% nationally, peaking at 11.7% in the Paris region.
  • Test Performance: The actual sensitivity for the 1.1 cut-off was 75.9%, significantly lower than manufacturer claims, though specificity remained near 100%.
  • Severity Metrics: The overall Infection Fatality Rate (IFR) was estimated at 0.8%, but this was heavily skewed by age, reaching 21.6% for those 80 and older. Similarly, the Infection Hospitalization Rate (IHR) for the 80+ demographic was 33.7%.

Limitations and Future Directions

While powerful, the model has limitations. The "infected" distribution was calibrated using RT-PCR positive individuals, who are typically symptomatic; asymptomatic individuals might have lower antibody levels, potentially leading to an underestimate of the non-responder rate. Additionally, the study focused on a specific 9-month window; applying this model to longer-term data would require incorporating antibody waning functions. Despite these factors, the continuous Bayesian approach provides a more robust framework for interpreting serological results than traditional binary classification.

Summary updated

Abstract

The individual results of SARS-CoV-2 serological tests measured after the first pandemic wave of 2020 cannot be directly interpreted as a probability of having been infected. Plus, these results are usually returned as a binary or ternary variable, relying on predefined cut-offs. We propose a Bayesian mixture model to estimate individual infection probabilities, based on 81,797 continuous anti-spike IgG tests from Euroimmun collected in France after the first wave. This approach used serological results as a continuous variable, and was therefore not based on diagnostic cut-offs. Cumulative incidence, which is necessary to compute infection probabilities, was estimated according to age and administrative region. In France, we found that a "negative" or a "positive" test, as classified by the manufacturer, could correspond to a probability of infection as high as 61.8% or as low as 67.7%, respectively. "Indeterminate" tests encompassed probabilities of infection ranging from 10.8 to 96.6%. Our model estimated tailored individual probabilities of SARS-CoV-2 infection based on age, region, and serological result. It can be applied in other contexts, if estimates of cumulative incidence are available.
BibTeX
@inproceedings{glemain2024estimating,
  author = {Benjamin Glémain and Xavier de Lamballerie and Marie Zins and Gianluca Severi and Mathilde Touvier and Jean‐François Deleuze and Fabrice Carrat and Pierre‐Yves Ancel and Marie-Aline Charles and Gianluca Severi and Mathilde Touvier and Marie Zins and Sofiane Kab and Adeline Renuy and Stéphane Le-Got and Céline Ribet and Mireille Pellicer and Emmanuel Wiernik and Marcel Goldberg and Fanny Artaud and Pascale Gerbouin-Rérolle and Mélody Enguix and C Laplanche and Roselyn Gomes-Rima and Lyan Hoang and Emmanuelle Correia and Alpha Amadou Barry and Nadège Senina and Julien Allègre and Fabien Szabo de Edelenyi and Nathalie Druesne‐Pecollo and Younes Esseddik and Serge Herçberg and Mélanie Deschasaux and Marie-Aline Charles and Valérie Benhammou and Anass Ritmi and Laetitia Marchand‐Martin and Cécile Zaros and Elodie Lordmi and Adriana Candea and Sophie de Visme and Thierry Siméon and Xavier Thierry and Bertrand Geay and Marie‐Noëlle Dufourg and Karen Milcent and Delphine Rahib and Nathalie Lydié and Clovis Lusivika‐Nzinga and Grégory Pannetier and Nathanaël Lapidus and Isabelle Goderel and Céline Dorival and Jérôme Nicol and Olivier Robineau and Cindy J. Lai and Liza Belhadji and Hélène Espérou and Sandrine Couffin-Cadièrgues and Jean‐Marie Gagliolo and Hélène Blanché and Jean-Marc Sébaoun and Jean-Christophe Beaudoin and Laëtitia Gressin and Valérie Deldrève and Ouissam Ouili and Jean‐François Deleuze and Läétitia Ninove and Stéphane Priet and Paola Mariela Saba Villarroel and Toscane Fourié and Souand Mohamed Ali and Abdenour Amroun and Morgan Seston and Nazlı Ayhan and Boris Pastorino and Xavier de Lamballerie and Nathanaël Lapidus and Fabrice Carrat and Fabrice Carrat},
  title = {{Estimating SARS-CoV-2 infection probabilities with serological data and a Bayesian mixture model}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-60060-3},
}
Figure