Machine Learning Reveals Three Distinct Antibody Journeys in Nursing Home Residents After COVID Vaccination

Older adults living in long-term care facilities were among the hardest hit populations throughout the COVID-19 pandemic, and they remain among the most immunologically complex. A new study published in Frontiers in Aging has now applied unsupervised machine learning to a year’s worth of blood samples from hundreds of Italian nursing home residents, and in doing so it has uncovered three fundamentally different antibody trajectories that conventional analysis had blurred into one. The findings offer a fresh, data-driven lens on how humoral immunity behaves in one of the most vulnerable groups in society.

The research, conducted within the framework of the COVAC-3 project, followed 388 residents aged 65 and older across 13 long-term care facilities in five Italian regions: Calabria, Friuli-Venezia Giulia, Lazio, Puglia, and Veneto. Participants were enrolled between November 2023 and February 2024 and tracked for 12 months, with blood drawn at four time points: baseline, three months, six months, and twelve months. At each visit, the researchers measured quantitative IgG antibodies against the SARS-CoV-2 Spike protein, the principal target of vaccine-induced immunity, alongside semi-quantitative antibodies against the Nucleocapsid protein, which serve as a serological fingerprint of prior natural infection.

The study population was typical of institutionalized elderly cohorts in its fragility. The mean age was 84.2 years, roughly 71 percent were women, and residents carried an average of four chronic conditions each, ranging from hypertension and cognitive decline to diabetes and heart failure. About half had received a booster dose during the 2023/2024 campaign, while 67 percent showed serological evidence of a previous SARS-CoV-2 infection. This overlapping mixture of repeated vaccination and natural exposure is precisely what makes immune responses in this population so difficult to disentangle using traditional statistical approaches that examine one variable at a time.

To cut through that complexity, the team turned to unsupervised machine learning. First, they used Principal Component Analysis to compress a large matrix of longitudinal antibody measurements, demographics, and vaccination history into a handful of dimensions capturing the main sources of variability. The first principal component was dominated by anti-Spike titers across all time points, essentially reflecting the magnitude and persistence of the vaccine-induced humoral response. The second was driven by anti-Nucleocapsid levels, a proxy for natural infection, and the third distinguished residents vaccinated within a month of baseline from those boosted earlier or not at all. Together, the first three components explained 66.4 percent of the total variance.

K-means clustering was then applied to those component scores, and the number of groups was chosen using a battery of internal validation metrics, including the Silhouette, Calinski-Harabasz, and Davies-Bouldin indices, combined with bootstrap resampling to test reproducibility. Three clusters emerged, and they proved remarkably stable: the mean Adjusted Rand Index across 100 bootstrap iterations was 0.74, indicating strong partition consistency. The three groups were also nearly equal in size, with 130, 125, and 133 residents respectively, and baseline characteristics differed significantly among them.

The first profile, labeled the Vaccine-Driven group, consisted of residents with moderate anti-Spike titers, little serological evidence of past infection, and a gradual, progressive antibody decline over the year. The second, the Hybrid Immunity group, comprised individuals with clear markers of prior infection but minimal uptake of the recent booster. Despite lower starting antibody concentrations, this group displayed the flattest trajectory of all, with antibody levels remaining essentially stable throughout the 12 months of follow-up. The third and most striking profile, the High-Peak Hybrid Responder group, combined recent boosting with prior infection and showed the highest baseline anti-Spike titers of the entire cohort, with a median of 10,700 BAU/mL, followed by the steepest rate of waning.

Linear mixed-effects models quantified these differences rigorously. The Vaccine-Driven reference cluster declined modestly but significantly over time, while the High-Peak Hybrid Responder group showed a significantly steeper slope, with an average marginal effect of minus 0.088 on log-transformed antibody levels per follow-up unit, compared with minus 0.031 in the reference group. The Hybrid Immunity group, by contrast, showed no statistically significant change at all. The pattern suggests that a high initial peak, driven by recent antigenic stimulation, is followed by a proportionally larger absolute drop, a phenomenon consistent with the natural decay kinetics observed after booster vaccination in older adults in previous studies.

Clinically, the picture was reassuring. Only 12 of the 388 residents, or 3.1 percent, developed documented symptomatic SARS-CoV-2 infection during the follow-up year, and there were no COVID-19-related deaths in the cohort. The few incident cases were spread across all three clusters, with no statistically significant association between cluster membership and infection risk, although the authors caution that the very low number of events limits the statistical power of that comparison. Notably, cognitive decline and dementia were more frequent in the High-Peak Hybrid Responder group, at nearly 70 percent, than in the other two profiles, hinting at possible links between immune trajectory patterns and clinical frailty dimensions that merit further investigation.

The researchers are careful to draw a clear line between description and prediction. Because the clustering itself incorporated longitudinal antibody measurements, the subsequent trajectory modeling serves as a descriptive characterization of the data-derived groups rather than an independent validation of clinical protection. Antibody titers, they note, represent only one dimension of immunity; they do not capture neutralization capacity or the T-cell responses that are particularly relevant in older adults. The observational design also meant booster timing could not be standardized, and a small number of post-baseline vaccinations and infections could not be fully accounted for in the primary model.

Even with those caveats, the study carries a meaningful message for public health. Rather than stratifying elderly populations by chronological age or comorbidity counts alone, multidimensional, data-driven profiling could help identify subgroups with genuinely different antibody dynamics, potentially informing more tailored booster schedules for long-term care residents. The work also demonstrates that unsupervised machine learning, applied thoughtfully to longitudinal serological data, can reveal structure in highly heterogeneous immune responses without imposing predefined biological assumptions. As booster campaigns continue and hybrid immunity becomes the norm rather than the exception, approaches like this one may become essential tools for understanding who retains protection, who wanes fastest, and when the next dose will do the most good.

Subject of Research: Longitudinal SARS-CoV-2 antibody response profiles in long-term care facility residents identified through unsupervised machine learning

Article Title: Longitudinal immune trajectories following SARS-CoV-2 vaccination in long-term care facility residents: an unsupervised analysis approach

Article References: Palmieri, A., Damiano, C., Leone, P., Trevisan, C., Malara, A., Zucchelli, A., Incalzi, R. A., Prato, R., Fortunato, F., Baldovin, T., Coin, A., Minchella, P., Giordano, S., Palamara, A. T., Onder, G., & Fedele, G. (2026). Longitudinal immune trajectories following SARS-CoV-2 vaccination in long-term care facility residents: an unsupervised analysis approach. Frontiers in Aging, 7, Article 1963384. https://doi.org/10.3389/fragi.2026.1963384

Image Credits: AI Generated

DOI: 10.3389/fragi.2026.1963384

Keywords: SARS-CoV-2, vaccination, long-term care facilities, machine learning, K-means clustering, principal component analysis, hybrid immunity, antibody waning, immunosenescence, older adults, humoral immunity, booster doses