More than one third of participants in two Swiss population-based cohort studies experienced a SARS-CoV-2 infection during the first Omicron wave, even though fewer than one in five reported a positive test. That gap between diagnosed and actual infections, documented by researchers analyzing data from the Zurich SARS-CoV-2 Cohort and the Zurich SARS-CoV-2 Vaccine Cohort, is the starting point for an exploratory study published in Diagnostic and Prognostic Research that set out to identify which variables best predict individual risk of Omicron infection.
Participants had been recruited between 2020 and 2021 and provided demographic information, vaccination histories, self-reported infections, and longitudinal serological data covering antibodies against the SARS-CoV-2 spike and nucleocapsid proteins, collected at six-month intervals. The analysis was restricted to participants with repeat blood samples, complete baseline demographic and behavioral data, and study visits before and after the outcome window.
The outcome was a SARS-CoV-2 infection during the first Omicron wave in Switzerland, from 1 January to 31 March 2022. Infections were counted either when participants self-reported a positive PCR or antigen test, or when serological measurements showed a seroconversion or at least a doubling of anti-S IgG, anti-S IgA, or anti-N IgG antibodies between two study visits, provided no vaccination could explain the rise. For infections identified only through antibody changes, the researchers estimated the most likely date of infection using a midpoint between visits, adjusted for the background incidence of reported cases in the canton.
The serological approach revealed substantial underdiagnosis. While 123 of 710 participants, or 17.3 percent, reported a positive test during the wave, 264 participants, or 37.2 percent, met the study’s infection criteria when antibody increases were included. The authors calculate that diagnosed infections alone underestimated the true infection burden by 53.4 percent. A similar pattern held for an extended period through the end of June 2022, where 22.5 percent had a documented positive test but 45.9 percent showed antibody evidence of infection. Most self-reported infections, 92.7 percent, were corroborated by corresponding antibody changes, supporting the coherence of the definition.
To identify predictors, the team fitted a series of logistic regression models and compared them using the Akaike Information Criterion, where a decrease of two or more points indicates improved fit, and assessed discrimination with the area under the receiver operating characteristic curve. All models included a core set of demographic and behavioral predictors: age, sex, smoking status, the presence of any of five comorbidities including diabetes, cancer, chronic kidney disease, respiratory disease, or hypertension, and a baseline compliance score reflecting social distancing, mask wearing, and hygiene behavior.
On top of that core, the researchers tested immunological and exposure-history variables one at a time. These included the most recent anti-N IgG or anti-S IgG antibody levels before the Omicron wave, whether a participant had received a booster vaccination in autumn 2021, the type of first exposure to the virus or vaccine, the total number of prior vaccinations and infections, and a categorical variable capturing the full sequence of infections and vaccinations. Applying backward stepwise selection to the complete predictor set produced the best-fitting model, with an AIC of 889.8 and an AUC of 0.69, a modest but measurable improvement over the base model’s AUC of 0.63.
The final model included age, sex, compliance with COVID-19 prevention guidelines, smoking status, comorbidities, prior anti-N IgG antibody levels, and the sequence of previous infections and vaccinations. Participants aged 65 and older had 50 to 60 percent lower odds of Omicron infection across all models, a finding consistent with other studies, though the authors note it remains unclear whether this reflects behavioral differences not captured by their compliance variable or immunological factors. Each tenfold increase in anti-N IgG levels before the wave was associated with roughly a 39.6 percent decrease in the odds of infection. Fewer prior exposures, whether through infection alone or infection followed by a single vaccination, were linked to higher odds of infection compared with a sequence of primary vaccination plus booster.
Behavioral predictors showed suggestive but statistically non-significant effects. Participants with low-to-moderate baseline compliance with COVID-19 regulations had 23 to 29 percent higher odds of infection than highly compliant participants, leading the authors to argue that behavior deserves a place in future models even though the association did not reach significance. The comorbidity indicator was not associated with infection odds, possibly because individual conditions were uncommon in this relatively small sample. An exploratory attempt to cluster participants by their longitudinal antibody trajectories using k-means clustering produced five distinct immune profiles but worsened model fit, suggesting that simple trajectory clustering did not add predictive value in this dataset.
The authors are careful to frame the work as exploratory. It is a secondary analysis of two existing cohorts with a fixed sample size and no prospective registration, and no formal sample size calculation was performed. Infection status relied on self-report and serological criteria rather than clinical adjudication, which may have introduced misclassification, and infection severity was not examined. Because the two cohorts sampled blood at different calendar times, some exposure-related bias across cohorts is possible, though sensitivity analyses using incidence-adjusted infection dates and an extended outcome window produced broadly consistent results. Multiple imputation for missing covariates, which increased the sample from 710 to 719, did not materially change the estimates, and modelling age as a continuous variable likewise left the other predictors stable.
The study’s contribution lies less in a ready-to-use clinical tool than in its evidence base for future model development. With an AUC of 0.69, discrimination remains modest, and the authors emphasize that calibration, clinical relevance, and validation in independent datasets will be essential before any prediction model reaches practice. Still, most predictors in the final model, apart from anti-N IgG, are readily available in clinical settings, and the sequence of infections and vaccinations emerged as a novel and practical predictor. The researchers suggest that penalized regression methods such as lasso or elastic net, richer longitudinal immune phenotyping, and validation in diverse populations with varying exposure histories are natural next steps toward models that could identify individuals at elevated risk of SARS-CoV-2 infection and inform targeted protective measures.
Subject of Research: Variable selection for prediction models of SARS-CoV-2 Omicron infection using two Swiss population-based cohort studies
Article Title: Variable selection for prediction models of Omicron infection: Insights from two population-based cohort studies
Article References: Lacy, J., Menges, D., Ballouz, T., Frei, A., Pellaton, C., Fenwick, C., Pantaleo, G., & Puhan, M. A. (2026). Variable selection for prediction models of Omicron infection: Insights from two population-based cohort studies. Diagnostic and Prognostic Research, 10(1), Article 14. https://doi.org/10.1186/s41512-026-00231-0
Image Credits: AI Generated
DOI: 10.1186/s41512-026-00231-0
Keywords: SARS-CoV-2, Omicron, prediction models, serology, anti-N IgG, vaccination sequence, cohort study, Zurich, AUC, variable selection, hybrid immunity, underdiagnosis
News Source: Kristina Jarvis. (October 29, 2026). Antibody Testing Reveals Hidden Omicron Infections in Swiss Cohort Studies. Scienmag.
